Automate, Build, or Buy The New Agency Playbook

Automate, Build, or Buy? The New Agency Playbook ![rw-book-cover](https://readwise-assets.s3.amazonaws.com/static/images/article1.be68295a7e40.png) 🔗 **Source:** [Automate, Build, or Buy? The New Agency Playbook](https://www.youtube.com/watch?v=Irj_t7mSOYU) > [!note]+ Document Notes > tags: [ai-agents, ai-automation, wordpress-development, agency-management, hosting-industry] > author: [MISSING] > source: Agency Action Podcast (The WP Minute Network) > reference: https://www.youtube.com/watch?v=Irj_t7mSOYU > date created: 2026-07-21 12:04 > > # 🏷️ Title: Automate, Build, or Buy? The New Agency Playbook >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - Matt (guest-hosting for Curt and Toby) talks with returning guest Brendan O'Connell about how agency AI workflows have shifted in the ~6 months since their last conversation. - Core arc: automation-first tooling (n8n) is being blended with, not replaced by, AI coding agents — and a new layer of always-on "personal agents" (Hermes, OpenClaw) is emerging for internal ops. - Key stakeholders: WordPress agency owners, plugin vendors (Gravity Forms as a case study), hosting companies (Hostinger named), and AI labs (Anthropic specifically) whose pricing/product tiers shape what agencies can afford to build. - Risks and tradeoffs: subsidized inference pricing won't last forever; utility plugins are exposed to AI-driven replacement while entrenched multi-integration products are more insulated; unclear whether agents can safely replace (vs. augment) project managers; local compute is currently too weak to run capable models, concentrating power with whoever controls hosted inference. - Bottom line: the agencies that win aren't the ones replacing WordPress with static-site tooling — they're the ones layering agents, dashboards, and automations *around* WordPress while WordPress itself closes the gap on MCP/Abilities API ergonomics. ## 📖 Main Content The episode is framed as a check-in: Brendan and Matt last recorded together roughly in January/February 2026, and both describe their AI workflows as materially different now. **From n8n to n8n-plus-AI.** Brendan still uses n8n heavily because it keeps infrastructure ownership legible — "you own sort of the infrastructure and you can kind of go in and tweak things a little more easily... it's not such a black box." What's changed is that AI now *writes* the n8n workflows, and automation is embedded inside AI-driven workflows rather than being a separate track. His worry: newer builders skip learning how APIs work and jump straight to "ask Claude to build a plugin," even when an existing automation tool (or plugin) already solves the problem — a classic buy-vs-build tension resurfacing at lower cost-to-build. **The subsidized-inference era and Anthropic's product ladder.** Matt lays out a reading of Anthropic's release cadence (Opus 4.6, then the "Mythos" mystique, then the time-limited "Fable" release repeatedly extended) as an attempt to find a price/product tier people will "softly land" in before costs escalate. Brendan's counter: open-weight models (he cites GLM 5.2) are competitive enough that they're likely reshaping that economic calculus in real time, which is why the trial windows keep extending. **Compute is the actual bottleneck.** Matt describes buying a high-powered laptop (RTX 4090, 24GB) specifically to run local models via LM Studio, and finding it still "dog slow" compared to hosted models. Brendan reports similar results with Google's Gemma model on a gaming laptop. Their shared conclusion: whoever controls affordable, fast inference compute — not model weights alone — wins the next phase, which is why hosting companies matter. **Hosting companies as the new "mall."** Matt's analogy: malls that didn't reinvent themselves (co-working, grocery, play centers) died; hosting companies facing WordPress-plugin commoditization may need to pivot from "we host websites" to "we host inference/agents." Brendan cites Hostinger's move into "OpenClaw" and "Hermes" hosting and one-click Docker deploys (including n8n) as an early version of that pivot. **Hermes as a daily-use agent.** Brendan runs Hermes (open source) on a VPS behind Tailscale for his company and personally via Discord/Slack/Telegram. It ingests Slack, Granola meeting notes, and Notion, and pushes daily updates — explicitly framed as augmentation ("I'm never in the business of replacing people... more like augmenting or trying to make sure they're not making mistakes") rather than replacing a project manager. He references a public Shopify engineering write-up about their internal agent "River," which improved by moving conversations out of DMs into public channels so the agent has richer context. **Matt's civic-tech experiment as a case study in AI value.** Matt describes a personal project that ingests YouTube transcripts of local government meetings, builds searchable/timestamped content, and now uses AI to generate blog posts and an AI-narrated (11 Labs) weekly podcast, published through WordPress with Wikipedia-style public correction for errors (the model gets names/local vocabulary wrong). A friend who edits the local paper reacted with "you're literally building a thing to replace my job" — Matt's stated framing is that the real unlock isn't replacing journalists but surfacing signal from volume no human has time to watch (the same logic he applies to cross-team meeting awareness inside an agency). **AI slop and the agent-to-agent future.** Both flag a stat (~40% of LinkedIn posts/comments cited as AI-generated) as evidence that naive "AI cold-outreach agents" mostly produce spam. Matt's hoped-for alternative: agents mediating contact instead of agents flooding inboxes directly ("stop hitting my inbox, just talk to my agent"). **WordPress's actual blocker: MCP/Abilities API friction.** Brendan wanted to connect a generic MCP server recently and found it more cumbersome than it should be, despite plugins (Meta Box, presumably Gravity Forms) shipping Abilities API support. Matt argues that once WordPress ships one-click/native MCP setup (core or a recommended plugin), the "static site builders are faster" narrative collapses — WordPress becomes equally conversational to build with, exposing whatever abilities each plugin registers. Matt already uses an application-password + local `.env` file to let Claude read/write WordPress content directly without full MCP tooling. Brendan's addition: this needs to work safely on a VPS (not just local CLI/SSH), and ideally live in WordPress core rather than solely as a paid hosting-company feature, so it doesn't become another monetized gate. ## 🗒️ Notes - Brendan's mental model shift: automation-only → automation *plus* AI, not automation *replaced by* AI. He still values n8n's transparency/ownership over "black box" agent behavior. - The "social experiment" framing (Matt's phrase) for LLM UX: fast models create impatience/ADHD-like task-switching; both hosts describe multi-tab parallel work sessions as a symptom of AI's "unlock" of latent creative drive in technical people. - Brendan explicitly worries about post-subsidy pricing changing behavior once the current cheap-token era ends — this is presented as a psychological dependency risk, not just a budgeting one. - Buy-vs-build keeps resurfacing: Facebook group anecdote (someone wants YouTube-to-WordPress integration; top suggestion is "have Claude build a plugin" vs. commenters pointing to existing tools like Flowmatic/n8n). - Gravity Forms as an insulation case study: a $200/year multi-integration product is harder to justify replacing with a custom AI-built form plugin than a "simple utility" plugin would be — network effects/add-on ecosystems create moat even in an AI-cheap-to-build era. - Composio mentioned as the connector layer that lets Hermes reach Slack/Notion/etc. without manual API key setup — lowers the barrier for non-technical agent setup. - Anthropic's own "half-baked Hermes"-style Slack agent release is cited as an example of a rushed competitive response that landed poorly. - Matt distinguishes his JeffBot (simple Cloudflare cron + RSS summarizer posting to Slack) from full agent infrastructure (Hermes/OpenClaw) — his stated preference is to wait for a native agent product from Anthropic/OpenAI rather than self-host agent infrastructure long-term. - Both explicitly reject "replace the project manager" framing in favor of "prevent context drift / catch dropped threads." - Troy Dean / "Viktor" is name-checked as an agency-facing agent tool Matt hasn't personally vetted, offered as a lead for listeners rather than an endorsement. ## 📚 References - [Agency Action Podcast](https://www.agencyactionpodcast.com) - [The WP Minute](https://thewpminute.com) - [Hostinger](https://www.hostinger.com) — cited for OpenClaw/Hermes hosting and Docker one-click deploys - [Compozio (Composio)](https://composio.dev) — cited for one-click API/service connections into Hermes - Shopify engineering blog post on their internal agent "River" (referenced verbally; specific URL not given in transcript — [MISSING]) - LM Studio — local LLM runner mentioned for on-device model testing ## 🗣️ Paraphrase Brendan explains that his workflow hasn't abandoned automation tools like n8n in favor of AI — instead, the two now work together, with AI often generating the n8n workflow itself, while n8n still gives him direct, inspectable control over infrastructure that a pure chat-agent black box wouldn't. He's more concerned that newer builders skip understanding how APIs and automation actually work, defaulting to "just have the AI build a plugin" even when a cheaper, already-built tool exists. Matt frames Anthropic's release pattern — a very capable Opus release, followed by scarcity messaging around a hypothetical ultra-premium "Mythos" tier, followed by a repeatedly-extended "Fable" trial — as a deliberate attempt to find the price point where users will settle, before costs climb. Brendan agrees but adds that competitive open-weight models are likely forcing Anthropic to keep adjusting that plan in real time, since users have a viable cheaper alternative waiting in the wings. ## 📰 Newsletter Agency owners still trying to decide between "automate it," "vibe-code it," or "just buy the plugin" will want this one. Brendan O'Connell joins Agency Action (Matt filling in for Curt and Toby) to talk through how his AI stack has evolved — n8n and AI coding agents now work together instead of one replacing the other, and he's running the open-source Hermes agent as an always-on ops layer across Slack, Notion, and Granola meeting notes. There's a sharp bit on hosting companies pivoting the way dying malls reinvented themselves (Hostinger's OpenClaw/Hermes hosting gets a shoutout), a candid worry about what happens to your workflow once today's subsidized AI pricing ends, and a genuinely useful WordPress-specific gripe: the MCP/Abilities API setup is still too clunky, and that friction — not model capability — is the real reason "just use Astro, it's faster" narratives keep getting airtime. Worth a listen if you're weighing whether to build your next client tool as a plugin, an automation, or an agent. ## ✅ Action Items #### DevRel - Developer experience: document a low-friction MCP + Abilities API setup path for WordPress site owners, closing the gap Brendan hit when connecting a generic MCP server. - Education: a workshop or guide comparing "buy an existing plugin/automation" vs. "have an AI agent build it," using the Facebook-group YouTube-embed anecdote as a worked example. #### WordPress - **Core:** prioritize native/one-click MCP setup so agent-driven site editing doesn't require manual SSH/env-file workarounds. - Core AI: track Abilities API adoption across major plugins (Gravity Forms, Meta Box) and publish a compatibility list for agency buyers. - Hosting: cross-host checklist for safely exposing MCP/agent access on a VPS (not just local dev), addressing Brendan's "not just in the hands of hosting companies" concern. #### Community - Ecosystem: solicit a follow-up from Shopify's "River" team or summarize their public write-up for the WordPress agency audience, since it's cited as directly relevant prior art. ## 🧪 Hypotheses 1. Plugins with narrow, single-purpose utility (e.g., simple contact forms) will see faster AI-driven substitution than plugins with deep multi-integration ecosystems (e.g., Gravity Forms). 2. Hosting companies that add agent/inference hosting (à la Hostinger's OpenClaw/Hermes offerings) will retain revenue better than those that stay purely web-hosting-focused as WordPress-only plugin sales decline. 3. Once WordPress ships native, low-friction MCP/Abilities API support, the perceived speed advantage of AI-native site builders (e.g., Astro-based stacks) over WordPress will largely disappear for common build tasks. 4. Agents given broad cross-channel context (Slack, meeting notes, project management tools) reduce reported "context drift" and missed cross-team handoffs more than agents restricted to a single data source. 5. As open-weight models (e.g., GLM 5.2) close the capability gap with frontier hosted models, commercial AI labs will be pressured to extend or expand discounted/trial access tiers rather than tighten them. ### 📊 Methodology - For (1): survey WordPress.org plugin categories, comparing year-over-year active-install trend for "utility" vs. "platform/ecosystem" plugins pre- and post- widespread AI-coding-agent adoption (roughly 2024 onward). - For (2): compare revenue-mix disclosures or public statements from hosting companies (Hostinger and comparable competitors) before and after launching agent-hosting products. - For (3): timed build-task benchmark — have equivalent testers build the same small site feature (form + page + CMS content) once WordPress ships native MCP support, in WordPress vs. an AI-native static-site tool, measuring time-to-completion. - For (4): controlled comparison of agent performance/accuracy on task handoff scenarios with single-source vs. multi-source (Slack + meeting notes + Notion) context feeds. - For (5): track publicly available token pricing and trial-tier duration changes from major labs against open-weight model benchmark releases over the same period. ### ❓ Questions - What specific friction points make MCP setup "cumbersome" today, and are they protocol-level or WordPress-plugin-implementation-level? - Is there quantitative data (beyond the LinkedIn 40%-AI-slop anecdote cited) on how much of current AI-agent cold-outreach actually converts, versus just adding noise? - How is Hermes's context-ingestion approach (Slack + Granola + Notion) governed with respect to sensitive data (Brendan notes email access was explicitly withheld)? - What would "core WordPress" ownership of agent/MCP tooling look like technically, given hosting companies have a commercial incentive to keep it as a paid add-on? ### 🔮 Future Research - Head-to-head cost/performance study of local (consumer GPU) vs. hosted inference for common agency AI tasks, to quantify how far local compute currently lags. - Longitudinal tracking of whether "agent-mediated contact" (Matt's proposed alternative to inbox spam) gains any real adoption versus remaining aspirational. - Case study of Matt's civic-tech project as a model for AI-assisted local journalism, including how the public-correction/wiki-style mechanism performs over time (accuracy trend of names/local vocabulary). ### 💡 Implications - **Practice:** agencies should audit their plugin stack for "utility vs. ecosystem" exposure now, since the two categories face very different AI-driven competitive pressure. - **Policy:** if agent-hosted infrastructure trends toward VPS/self-hosted setups (Tailscale-secured, per Brendan's example), agencies will need updated security guidance for client-facing agent deployments — this isn't covered by traditional WordPress hosting security checklists. - **Education:** the buy-vs-build decision framework needs updating for an era where "build" costs have collapsed; standard total-cost-of-ownership models likely underweight integration/maintenance costs relative to a $200/year finished product. - **Funding models:** if inference-hosting becomes hosting companies' primary margin driver (the "mall" pivot), pricing models for WordPress hosting overall may shift, affecting agencies that resell hosting to clients. ### 🧠 Learning Styles Assessment - Reading level estimate: this is a lightly-cleaned conversational transcript (filler words retained, no punctuation-perfect sentence boundaries); a Flesch-Kincaid-style estimate lands around grade 8–9 for the underlying ideas, though the raw transcript reads rougher due to spoken-language artifacts. #### 👀 Visual learners - Present: none — pure audio/transcript content, no diagrams or visuals referenced. - Improve: an architecture diagram of the Hermes context pipeline (Slack + Granola + Notion → agent → Slack/Notion updates) would clarify the most technical segment of the episode. #### 🔊 Aural learners - Present: this *is* the native format — a two-person conversational podcast. - Improve: a chaptered audio version with spoken segment markers would help, since the current transcript has no timestamps to navigate by. #### 🗣️ Verbal learners - Present: strong use of analogy (hosting-companies-as-malls) and concrete examples (Gravity Forms pricing, RTX 4090 test) to explain abstract claims. - Improve: a short glossary of terms used loosely (MCP, Abilities API, Hermes, OpenClaw) would help listeners unfamiliar with the WordPress-AI ecosystem. #### 🧩 Kinesthetic learners - Present: none directly — no hands-on steps given. - Improve: a companion checklist for "set up Hermes with Composio in under an hour" (referenced verbally as achievable) would convert the anecdote into a reproducible exercise. #### 🧠 Logical learners - Present: the buy-vs-build and automate-vs-agent framings function as implicit decision trees throughout. - Improve: an explicit decision matrix (utility plugin vs. ecosystem plugin; local vs. hosted inference; augment vs. replace project manager) would make the episode's recurring logic reusable. #### 🧑‍🤝‍🧑 Social learners - Present: the episode itself is a peer-to-peer dialogue model well-suited to discussion-style consumption. - Improve: WP Minute's member Slack (mentioned at the close) is a built-in discussion venue — linking specific claims (e.g., the Anthropic pricing-ladder theory) to a discussion thread would extend this naturally. #### 🧘‍♀️ Solitary learners - Present: none explicit. - Improve: a self-check prompt ("audit your own plugin stack for utility vs. ecosystem exposure") would turn passive listening into reflection. #### 🌿 Naturalistic learners - Present: the mall-reinvention analogy implicitly treats hosting-company business models as an ecosystem under environmental pressure. - Improve: extending that analogy explicitly (what's the "invasive species" here — open-weight models? agent hosting?) would sharpen it for this learner type. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: N/A to the source video itself; this note uses standard heading hierarchy. - Improve: original YouTube video should carry a proper title and description with speaker names for screen-reader users browsing search/reader views. #### Keyboard navigation - Present: N/A (audio/video content). - Improve: if published as a written companion post, ensure skip-links and logical heading order (as done here). #### Contrast and readability - Present: N/A — no visual UI discussed. - Improve: N/A. #### Text resizing and reflow - Present: N/A. - Improve: N/A. #### Media and captions - Present: a transcript exists (this document's source), which is the main accessibility asset for a podcast. - Improve: the transcript lacks timestamps and speaker labels beyond "Matt"/implied Brendan turns — adding both would materially help Deaf/hard-of-hearing users and anyone navigating by topic. #### Forms and interactions - Present: N/A — no interactive elements in the source. - Improve: N/A. #### Standards reference - Primary WCAG 2.2 AA gap: missing synchronized captions/timestamps and speaker identification in the transcript (relates to 1.2.x Time-based Media criteria). ### 🌍 DEIB Assessment #### Cultural breadth - Present: discussion is US-centric (local government/local paper example, US mall references); no international WordPress-agency perspectives included. - Improve: a follow-up episode with non-US agency owners would test whether the "subsidy era ending" and hosting-pivot theses hold outside US pricing/market conditions. #### Inclusive language - Present: informal, jargon-heavy (MCP, abilities API, "vibe-code") without definitions. - Improve: a plain-language glossary would make this accessible to non-technical agency stakeholders (sales, ops) rather than developers only. #### Neurodiversity support - Present: conversational, non-linear structure typical of unscripted podcasts; topics loop back on themselves (e.g., returning to MCP friction at the end after touching it mid-episode). - Improve: chaptering by topic (rather than strict chronological order) would help listeners who find topic-jumping disorienting. #### Representation - Present: two participants, both apparently male, both technical/developer-role WordPress professionals; no other roles (designers, PMs, non-technical founders) represented directly. - Improve: future episodes could include a project-manager or client-side perspective on the "agent replacing PM" question raised but not resolved here. #### Psychological safety - Present: respectful, collegial tone throughout; disagreement (e.g., on open-weight models' competitive impact) is framed as building on each other's points rather than adversarial. - Improve: N/A — no concerns identified. ### ✅ Logical Validity #### Strong reasoning - **Analogical reasoning:** the hosting-companies-as-malls analogy (mid-episode) is a strong, well-developed comparison — it maps a specific structural pattern (diversify offerings or die) onto an unfamiliar domain (hosting economics) in a way that generates a testable prediction (agent/inference hosting as the "grocery store in the mall"). - **Causal reasoning:** Brendan's claim that open-weight model competitiveness (GLM 5.2) is forcing Anthropic to repeatedly extend the Fable trial is a plausible causal link supported by observable behavior (repeated extensions) rather than asserted without evidence. #### Fallacies - **Hasty generalization:** Matt's claim that "40% of LinkedIn posts... are just AI slop" is cited as "a survey... some outlet did" without naming the source or methodology — presented with more confidence than the sourcing supports. Fix: name the source and sample size before using it as evidence for a broader claim about communication's future. - **Unfalsifiable framing:** the theory that Anthropic is deliberately engineering a "stair-step" pricing conditioning strategy is plausible but stated as near-certain intent ("that's how they're going to condition us") without acknowledging simpler alternative explanations (e.g., ordinary iterative product releases, safety-driven staged rollouts). Fix: flag it explicitly as a hypothesis about corporate intent rather than an established fact. #### Tone of voice - Conversational, informal, collegial — frequent hedging ("I don't know," "I guess") signals genuine uncertainty rather than overconfidence, which is a mitigating factor against the fallacies above. ### 🛡️ DARVO Analysis - **Deny:** not applicable — no accusation is being responded to in this conversation. - **Attack:** not applicable — no party's credibility is under attack; disagreements (e.g., over open-weight model competitiveness) are collaborative, not adversarial. - **Reverse victim/offender:** not applicable. - Overall: this is a cooperative dialogue between peers with no accusation-response dynamic present; DARVO analysis does not apply to this content. ## 🎯 Conclusion The episode's throughline is that AI hasn't eliminated the older WordPress-agency toolkit (automation platforms, established plugins, hosting relationships) — it's layered on top of it, shifting where value and risk sit. Utility plugins and pure automation-only workflows face real substitution pressure; ecosystem-anchored products, hosting providers willing to pivot into agent/inference hosting, and WordPress itself (once MCP/Abilities API friction drops) are positioned to absorb that shift rather than be replaced by it. Practical next step for listeners: audit your own plugin stack and client-agent setups against the utility-vs-ecosystem and augment-vs-replace framings raised here, rather than assuming either "AI replaces this" or "AI changes nothing." ## 🔄 Refactor Appendix — Automate, Build, or Buy? The New Agency Playbook > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). No Readwise ingestion ID exists for this source (raw pasted transcript, not a Reader import) — quote blocks below are verbatim transcript excerpts without `ID:rw_` markers, since fabricating IDs would misrepresent provenance. ### 📚 Literature Atoms (H3) #### Utility plugins face more AI substitution risk than ecosystem plugins > [!quote] > "if you have just like utility products um those will probably the be the first to go... if I need something to replicate on all my client sites with all the add-ons and connections. Do I want to risk that for 200 bucks a year to buy Gravity Forms? Probably not." **Context & use**: - Mid-episode, during the WordPress-product-risk discussion - Directly actionable for agency owners deciding which client plugins to defend vs. treat as replaceable **Tags**: #xref  #### Open-weight model competitiveness is reshaping frontier-lab pricing tiers > [!quote] > "we have these open weight model like GLM... very capable. And so... they're seeing these open weight models come from behind thinking that that's changing their their calculus" **Context & use**: - Discussion of Anthropic's Opus/Mythos/Fable tiering strategy - Supports hypothesis 5 above; testable via public pricing/trial-duration tracking **Tags**: #xref  #### Hosting companies are pivoting from web hosting to agent/inference hosting > [!quote] > "Hostinger... They've kind of embraced... the AI agent craze... they were one of the first to really offer open claw hosting and now Hermes hosting" **Context & use**: - Follows the "hosting companies as malls" analogy - Relevant to agency owners choosing hosting partners for client agent deployments **Tags**: #xref #addendum  ### 🧠 Permanent Candidates (H4) #### Buy-vs-build economics inverted, not eliminated, by cheap AI coding **Claim** (one sentence, Courtney's voice): Cheap AI-assisted coding lowers the cost of "build" enough to change the buy-vs-build calculus, but it doesn't erase the moat that multi-integration, ecosystem-anchored products hold over single-purpose utilities. **Evidence pointer**: [[#Utility plugins face more AI substitution risk than ecosystem plugins]] **Scope & conditions**: - Holds when: the existing product has deep integration/add-on surface area (Gravity Forms example) that a custom build would need to replicate. - Breaks when: the utility is genuinely narrow (a single contact form, a single API pull) with no meaningful ecosystem lock-in. **Links**: - [[Hosting-as-inference-infrastructure]] — because contrasts on which layer of the stack captures value as building gets cheaper  ### ✏️ Fleeting Queue - [F1] Find and read Shopify's public engineering write-up on their internal agent "River" (public-channel-context strategy) — referenced verbally, URL not given in source #xref - [F2] Look into Troy Dean / "Viktor" as an agency-specific agent tool — mentioned as popular in agency circles but not independently verified in this episode #xref ### → Things 3 (max 2 actions) - [ ] Audit current WordPress client plugin stack for utility-vs-ecosystem AI-substitution exposure — agency risk review #things --- ## Full Transcript [00:01] We're back with another agency I'm Wait [00:03] a minute. No, I'm filling in the roles. [00:06] Two four very big shoes. Curt and Toby [00:11] are on vacation, so I'm covering for [00:13] them on the Agency Action podcast. So, [00:16] welcome everyone. Uh it's your lovely [00:19] host Matt with special guest Brendan [00:23] O'Connell. We're going to talk all about [00:25] AI stuff today. [00:27] Brendan and I chatted uh we've chatted [00:29] multiple times on on the WP Minute [00:31] podcast, but this is going to be maybe [00:33] about a 6-month hiatus since you and I [00:36] have officially hit the airwaves, [00:37] Brendan. What do you think? [00:40] >> Yeah, I think back in January or [00:42] February. [00:43] So, good to be back. [00:44] >> Lots of things have changed uh obviously [00:47] in the technology um of uh AI and [00:50] website building and agency life. Um [00:53] Brendan has done a bunch in his personal [00:56] career, professional career. Uh so, [00:58] excited to talk about how he has seen [01:00] the landscape shift since uh he and I [01:04] last chatted. And maybe it resonates [01:06] with you, the agency owner who's like [01:08] listening to all this news of uh [01:12] you know, hey, we're going to lose our [01:13] jobs. Uh is is WordPress in trouble? [01:15] This stuff we were questioning like a [01:17] year ago [01:18] maybe being solidified a little bit more [01:20] on the product side these days in terms [01:22] of that uh in the headlines, but really [01:24] interested to dive in. I mean, I'm using [01:25] AI differently than I was just 6 months [01:28] ago. Um and I just want to unpack a lot [01:31] of that stuff today. So, Agency Action [01:32] podcast uh agencyactionpodcast.com [01:35] where you can catch up [01:37] with all of these podcasts uh episodes [01:40] um via a retro arcade machine that I [01:45] built with AI as a way to play a video [01:49] game of Agency Action. Uh if you don't [01:51] want to listen to the whole episode. [01:52] It's actually one of the ways that I [01:53] experience Agency Action. So, check that [01:55] out. Agency Action Podcast .com. [01:58] Brandon, I want to frame this question [02:00] for you. Um [02:03] when AI started to really become [02:07] more mainstream for us prosumers and [02:10] professionals in the space, [02:12] you were doing this at you were doing a [02:14] lot with like n8n and automation at that [02:17] time. Like you were really big into that [02:19] and then like AI started coming in. [02:22] Where have you landed with like the days [02:25] of [02:26] Brandon, I'm going to create these [02:27] blueprints of n8n to [02:30] now like, oh, maybe AI is just the mind [02:33] hive that we need to do all this stuff. [02:36] Where have you landed in your technology [02:38] stack these days? A traditional [02:40] automation versus AI. [02:44] >> Yeah, I mean that that's a that's a good [02:45] question cuz I think [02:47] for a lot of people [02:48] AI is still kind of a black box, right? [02:50] Because they they put they put a prompt [02:52] in and it does something and they it [02:55] spits something out and n8n is still a [02:57] visual workflow builder and [02:59] you know, I I actually do still I do I [03:01] do still use n8n quite a bit because it [03:03] is you know, you own sort of the [03:05] infrastructure and you can kind of go in [03:07] and and tweak things a little more [03:09] easily and [03:10] uh [03:11] you know, it's not such a black box, [03:12] right? So, uh you know, I do still use [03:15] it quite a bit actually and I use them [03:17] in conjunction, right? Cuz you can get [03:19] you can get your AI to to build your n8n [03:21] workflow for you uh these days, which is [03:23] pretty nice. [03:25] But, you know, I also use I use [03:29] automation within within my AI workflows [03:32] as well. Um [03:33] so, I don't know. I mean it's a [03:36] it's an interesting concept, right? That [03:38] now it's before it was right it was just [03:41] sort of build your build your [03:42] automation, maybe put a little AI in [03:44] there and now it's maybe a little [03:46] heavier on the AI side. [03:49] Um but I think people are probably a [03:50] little [03:52] inclined to skip the automation step and [03:54] not really understand how APIs work and [03:57] I think that's still a valuable skill. [04:01] >> Yeah. And and I think that's one of the [04:04] one of the unlocks for agency owners who [04:06] might be building [04:08] um [04:10] stuff alongside of their, you know, [04:13] client WordPress sites [04:14] because I would say that the [04:18] benefit to uh [04:22] AI right now, there's obviously [04:24] [clears throat] a lot of people like, [04:25] "Hey, I'm just going to like replace [04:26] WordPress with Astro or some other like [04:29] static thing." That's not really the [04:31] conversation I want to have. I want to [04:32] have the conversation of like, "Who's [04:34] still building with WordPress [04:36] but expanding their [04:40] value to the customer by building things [04:43] that sit alongside WordPress, right? [04:46] Reporting tools, [04:48] um you know, dashboards, bringing in [04:50] data and other content, you know, into [04:53] WordPress, connecting to other APIs that [04:56] you otherwise might not be you might [04:58] have a plugin for or something like [04:59] that. I'm really interested to see how [05:01] this evolves over the next I mean, I'm [05:03] just going to call it the 6 months, [05:05] right? For the all the way throughout [05:06] the end of the year [05:08] and see what people are building as [05:09] agency owners that ride alongside [05:12] WordPress instead of just replace [05:14] WordPress. That's not really the flag I [05:15] want to fly, uh if that makes sense. [05:18] >> [laughter] [05:19] >> Yeah, [05:20] I mean [05:21] >> even just today I was in some Facebook [05:24] group and someone was trying to [05:26] you know, ask about how do I get a bunch [05:28] of YouTube videos into my WordPress site [05:32] or something like that and you know, [05:34] most of the suggestions were [05:36] oh, well, just ask Claude to build a [05:38] plugin to do it. And [05:41] you know, some of the [05:42] that was most of the suggestions and [05:44] it's like, well, yeah, I mean, you I [05:46] guess you could do that and that's [05:48] probably how most of people are going to [05:50] approach it these days. Um but it's like [05:53] the most a lot of those comments at the [05:55] bottom were like use N8N or use another [05:58] automation tool like Flowmatic or some [05:59] some other WordPress plugin that already [06:01] exists that that does that for you. Uh [06:04] so I mean yeah, there's you know, it's [06:06] the it's kind of the age-old question of [06:08] like buy versus build. [06:10] Uh and you know, the the the [06:14] the idea of you know, what what it takes [06:16] to build something is going down and [06:18] down and down. So uh yeah, you could you [06:21] could just buy build something now for [06:24] sure, but you know, there's solutions [06:26] out there that already exist and [06:28] understanding how stuff works is still [06:30] it's still important. [06:32] >> Yeah. [06:33] How how have you adjusted to [06:36] the speed, I'm going to put that in air [06:39] quotes, the speed of AI? [06:42] Uh [06:42] if you can, you know, reflect maybe even [06:44] a year, two years ago, how have you [06:47] adjusted to the speed of waiting, maybe [06:50] that's not the maybe speed isn't the [06:51] right word, but [06:53] how have you adjusted to waiting for [06:54] something to finish in AI? Do you feel [06:58] yourself getting like more antsy, like [07:00] come on now, finish this, you know, like [07:02] finish this task when like a year ago [07:04] you'd be like, I'll wait an hour for [07:05] this thing to finish cuz now you could [07:06] never do it. Do you feel yourself [07:08] shifting to like if this thing's not [07:10] done in 5 minutes, I need to go find [07:11] like another LLM to to deal with. Do you [07:14] feel like that psychosis or that emotion [07:18] changing in you as you've become a more [07:21] maybe AI forward developer over the last [07:23] two, three, one to two, three years? [07:27] >> Yeah, a little bit. I mean, I think I [07:29] think this whole [07:31] sort of social experiment that they've [07:33] been running on us [07:34] >> Yes, that's where I was going. Thank [07:36] you. [07:36] >> [laughter] [07:37] >> Yeah, this is this has turned us all [07:38] into uh uh incredible people that have [07:42] uh you know, we've we've been building a [07:44] lot more, right? But we've also [07:47] our ADHD is off the charts now. [07:50] When I switch to [07:52] a lower a lower tier model, I've been [07:54] using command code a lot. I don't know [07:56] if you've checked that out. [07:58] But it lets you kind of switch between [07:59] models on command. Uh and you know, so [08:03] so I'll sometimes switch between some of [08:05] the open weight models and and try those [08:06] out. And man, some of those are are [08:09] really nice, right? Cuz they're super [08:10] cheap, but they are a lot slower than [08:12] Claude. And yeah, like I'll just sit [08:14] there waiting for a long time. [08:16] And then I'll end up going back to [08:17] Claude a lot. [08:19] Switch back to Codex or whatever, but [08:21] yeah, you know, and then we open up [08:23] another terminal and and start working [08:25] on another task or use the the /btw [08:27] feature and and open up another chat [08:30] that way. [08:31] Um [08:32] Yeah, it's it's hard to resist that, [08:34] right? Because we feel like our our time [08:36] is so precious now and we need to always [08:39] be productive and be working to get get [08:42] the most out of that 5-hour window and [08:44] and get all of our tokens in and get the [08:46] most out of our subscription. [08:48] Um [08:50] I try to resist it the most I can, [08:51] though, because it's it's not good for [08:54] our brains. [08:56] >> Yeah, I mean I I I can definitely feel [08:57] that. I wonder if it's just because we [08:59] are [09:01] whether we realize it or not, like we [09:03] are creative. I'm just speaking, I [09:05] guess, for myself, but I imagine seeing [09:07] what you do, what you post online and [09:09] stuff, we're we're kind of similar is [09:11] we're creative people. We like to build [09:13] things and we like we like to have these [09:16] ideas come to life. [09:18] Therefore, like this is crack to us. [09:22] >> [laughter] [09:22] >> All right? Like this is oh my god. I've [09:25] had a gazillion ideas and now I can just [09:29] open up a tab. One idea's cooking, move [09:32] to the next, that idea's cooking. Got a [09:35] third idea? [09:36] Do it over here, and I'll often catch [09:39] myself [09:41] is in in any given what I'll call work [09:44] session, I'll have four tabs. [09:47] And this is probably I'm probably, you [09:48] know, not running as much as most [09:50] people, but like four tabs of four [09:52] different things being built, and I'm [09:54] just jumping in between them. Meanwhile, [09:57] I'll have like a chat GPT window open in [10:00] my browser [10:02] asking it questions, going like, "Hey, I [10:03] got this other idea, like what do you [10:05] like?" Just like I'm just throwing [10:07] spaghetti against the wall over here [10:09] while cooking in the kitchen with these [10:11] four apps, and at the end of the day, [10:14] like there is a sense of accomplishment, [10:17] but it's also like, "Holy now what [10:20] do I do with all of this stuff?" But I [10:22] I think that's what they unlocked that [10:23] or they hit us in a different way that [10:26] no other product has before as creative [10:28] people. And that's like the thing. Like [10:30] that's where the pressure comes from for [10:32] me. I don't find myself worrying about [10:35] the cost or the tokens and stuff like [10:37] that. It's more about like, "Wow, I can [10:39] get all these ideas to life." I I don't [10:41] know if you feel the same way. [10:43] >> Yeah, well, and and and honestly, like I [10:45] never really ever felt like a very [10:47] creative person to begin with. Like I [10:49] never felt like an idea person or or [10:52] someone who was, [10:53] you know, coming up with the best next [10:55] thing or whatever. And and I probably [10:57] still am not, obviously. I'm not coming [10:59] up with the next million-dollar thing, [11:00] right? But [11:01] yeah, I mean, I think you're right that [11:02] they've they've [11:03] figured out how to unlock that in in [11:06] people that work with code and work with [11:08] apps and things like that and work with [11:09] websites. And they've it's it's [11:12] addicting, right? You're absolutely [11:13] right. And and I and I do kind of worry [11:16] that when we reach the end of the [11:18] subsidy era, what that's going to do [11:22] because right now, you know, I I I [11:25] I used [11:26] this thing called Agent View that shows [11:28] like how how much [11:30] how much we've spent, right? [11:33] Uh in terms of our token usage, and [11:35] we're getting a good deal right now [11:37] because uh it's subsidized. And when [11:40] that era ends, if it when if and when it [11:42] ends, um [11:44] it's going to it's going to be hard [11:46] because we're we're getting used to this [11:48] feeling of of creative juices flowing [11:50] all the time, and [11:51] yeah, it's going to be it's going to be [11:52] interesting when that happens. [11:55] >> I'm really concerned [11:57] uh [11:58] about that, and I think [12:01] you know, I I I think what you see with [12:03] Claude uh or excuse me, I should say [12:05] Anthropic. What you What you see with [12:07] Anthropic [12:08] is this balance between [12:12] I I feel like when Opus 4.6 hit, [12:15] they [12:17] did that as like um [12:20] you know, uh [12:21] I don't know what the proper phrase [12:22] would be, but they're like, "Hey, we got [12:23] this awesome thing that we have, you [12:25] know, internally. Let's let it loose in [12:28] in Opus 4.6. [12:30] Get people to be like, "Oh my god, this [12:32] is [12:33] like this is the like coding model, at [12:35] least for coding stuff. [12:37] This is the model." And then they [12:38] realized, "Oh, maybe we gave them too [12:40] much. Maybe 4.6 like we gave them too [12:43] much with 4.6, and ever since 4.6, [12:47] they've been trying to rebalance what [12:50] that product matrix is going to look [12:51] like. We hear things about Mythos, which [12:55] they've already started to condition us [12:56] like, "You'll never have that. And if [12:58] you do, it's going to be like $20,000 [13:00] per million tokens." [13:02] Um [13:04] But but then they released Fable, and [13:06] that was like, "Oh, the government had [13:07] to step in." And then they said, "Okay, [13:09] we have it for 7 days, and then 14 [13:11] days." And it just got re-extended again [13:13] this weekend at the time of this [13:14] recording. [13:15] And I feel like they're trying to [13:16] balance that product matrix of [13:21] how do we get people to sort of softly [13:23] land in like a middle area that they can [13:28] that we can charge them a couple hundred [13:29] bucks with and that'll be the floor. [13:33] But then as soon as you start to look to [13:35] that fable or that mythos, now you're [13:37] talking thousands of dollars, right? And [13:40] that's how they're going to condition us [13:41] to sort of stair-step [13:44] um until eventually you're like, "Oh [13:46] god, do I can I afford this?" One. And [13:49] two, like, "Do I just want to wait more [13:51] time and use Sonnet 5 or Sonnet 7 in the [13:55] future?" That's going to cost like 50 [13:58] bucks a month, but take 10 times longer [14:00] to produce something. Um I I feel like [14:03] out of all of them, they're the ones [14:04] that are trying to really balance that [14:06] matrix of [14:08] not rug pull, but I guess condition us [14:10] to what the future costs are going to [14:12] be. [14:12] >> time though, we have these open weight [14:14] model like GLM [14:16] GLM 5.2 is coming up behind them and is [14:19] very capable. And so, I think they I [14:23] think you're right that that's their [14:25] their thinking, but then [14:27] they're they're seeing these open weight [14:28] models come from behind thinking [14:31] that that's changing their their [14:32] calculus, right? So, I think that [14:34] they're That's why they they keep they [14:36] they keep having to kind of extend these [14:38] these [14:39] these trials for fable. They keep having [14:42] to release newer newer models. I think [14:43] it's changing their their economics a [14:46] lot. Uh and so, [14:49] you know, I I [14:50] Who knows what's going to happen because [14:51] I think that that was that was probably [14:53] their thinking, right? That they Oh, [14:54] we'll just [14:55] >> we'll just give them these, you know, [14:56] 4.6 level models and call it a day and [14:58] then keep these you know, lock these [15:01] higher level models behind these [15:02] paywalls, but I don't think that's going [15:04] to be the case anymore because these [15:05] open weight models are are very [15:07] competitive now. [15:09] >> Yeah, the [15:12] I guess the the the big challenge with [15:13] that is cuz I've tried [15:15] um [15:17] So, I don't use command code, but I've [15:18] tested things with uh, open router. I [15:21] believe that's what it's called. Open [15:22] router. I haven't spent much time, but [15:23] I've just like experimented with some [15:24] things. And then I tried, um, [15:27] another local tool which I bought, uh, [15:32] one night. I was like, "Oh, I'm going to [15:33] run." Um, let me just see if I can find [15:35] the app. I always forget the name of it [15:37] cuz I don't use it often. Let me just [15:39] see if I see it. LM Studio, which allows [15:42] you to run [15:44] local models on your desktop. [15:45] [clears throat] [15:47] Um, [15:48] and one night I I bought, uh, [15:50] this high-powered [15:52] Windows laptop cuz it has the Nvidia [15:55] 4090 graphics card in it. So, I was [15:57] like, "Oh, I'll [15:59] business expense." [16:00] >> [laughter] [16:01] >> Stupid. Uh, but anyway, so I I bought [16:04] it, uh, just to run these models. And [16:06] even with that, uh, even with that [16:08] graphics card, which has 24 GB, [16:11] I mean, obviously it's not a desktop, [16:12] which has 32, uh, for the graphics card, [16:14] but even that was [16:16] more powerful than [16:18] what I'm going to run on my MacBook in [16:19] terms of a dedicated, uh, graphics card, [16:23] GPU. And, um, [16:25] running those models locally is still [16:29] dog slow. [16:31] And I think where these guys are going [16:32] to win, [16:34] um, [16:35] is how do you manage this compute power [16:37] to run [16:39] these models? And and who's going to win [16:41] at that? And and I know a lot of people [16:44] talk about this and like data centers [16:45] and and all that stuff. And Google [16:48] versus Nvidia, uh, you know, versus Open [16:51] AI and Apple. Like, who's going to build [16:53] all this processing? [16:54] Um, I want open source to win. My My [16:57] fear is [17:00] that we're not going to have enough [17:02] CPU power to run this stuff until these [17:04] things get really, really much smaller [17:06] to run on on local machines. [17:09] Um, [17:10] that's my fear, anyway. [17:14] >> Yeah. Yeah, no, I mean, for sure. I I [17:16] mean, I to run even just like uh, base [17:18] level Google Gemma 4 model and then my [17:21] my apparently powerful gaming laptop can [17:24] barely handle it, too. So, you're right [17:27] on that front. [17:28] You know, I think give it give it a year [17:30] and we'll see these we'll see these be [17:32] able to run on a phone, maybe. Um [17:35] you know, but at the same time yeah, I [17:37] mean like something like command code I [17:38] guess is still like I'm accessing these [17:40] open way models not on my local device, [17:42] right? I'm I'm hitting some server that [17:44] they that they're running. [17:45] Uh [17:46] and so I'm you know, they're probably [17:48] not using the best compute power that [17:50] Anthropic's using, but so yeah, I don't [17:53] know how that compute matrix is working [17:55] there, but [17:57] yeah, I don't know how that's going to [17:59] unfold. So. [18:00] >> Yeah, the future is definitely [18:01] interesting for these for like hosting [18:03] companies. [18:04] Um [18:06] Yeah, and what they do because we think [18:09] of hosting in the traditional sense like [18:10] to pull it back to WordPress. Like [18:13] obviously [18:15] if you start seeing product companies in [18:16] WordPress starting to um [18:20] like decline in sales. I mean I guess [18:22] the question is is like what kind of [18:23] product do you have? And you know, how [18:26] are you you know, if you have just like [18:28] utility products [18:30] um [18:31] those will probably the be the first to [18:33] go, right? If you have small little [18:35] plugins that do things that you can [18:37] quickly whip up with a Codex or a [18:39] Claude, like yeah, it's [18:42] going to be a tough business. Um [18:44] So, like what can you do to extend that? [18:46] Um [clears throat] [18:47] folks like in my [18:50] career, my thing that puts the the food [18:52] on the table for me, Gravity Forms, like [18:54] we have a little bit more insulation [18:56] against that because it's a big plugin. [18:59] When you start to [19:01] look at it like am I going to code a [19:02] replacement for this? Yeah, maybe I'll [19:04] code a [19:06] simple contact form replacement, but if [19:08] I need something to replicate on all my [19:10] client sites with all the add-ons and [19:12] connections. Do I want to risk that for [19:14] 200 bucks a year to buy Gravity Forms? [19:16] Probably not. Let me just spend the 200 [19:18] bucks. [19:19] Um [19:21] but the [19:23] uh hosting companies [19:25] would obviously be next in line, right? [19:27] If If If we all are having this impact [19:29] in WordPress with WordPress getting um [19:32] pulled [19:34] away from because of AI, hosting [19:36] companies are going to have to change. [19:37] And I was thinking the other day [19:40] hosting companies might be like [19:42] your mall. Now, I don't know about your [19:44] mall if you have a mall in your area in [19:47] the US anyway. It's the places where [19:49] people used to go to the one big [19:51] building with a bunch of stores in it. [19:53] And now nobody goes to one big building [19:55] with a bunch of stores in it anymore. [19:57] Like I don't know anyone who goes [19:59] shopping at the mall like it used to. [20:00] So, I see a few malls in my area who are [20:03] like reinventing themselves. Now they [20:05] have grocery stores in them. Now they [20:07] have co-working spaces in them. Now they [20:09] have play centers in them. It's not like [20:12] 25 retail outlets anymore. [20:14] >> Right. [20:15] >> Hosting companies might have to shift to [20:17] be like, "Okay, we don't do websites [20:18] anymore. We do [20:20] inference. We do all of this like [20:24] open router compute." Like they have to [20:26] kind of shift their model into [20:28] you know, the the future world. [20:31] >> Yeah, I mean, I think [20:32] there's there's a couple that come I [20:34] mean, like Hostinger, I guess, is is one [20:36] that comes off the top of my head. [20:38] They've kind of embraced, I think, the [20:41] AI agent [20:43] craze, right? They They were one of the [20:45] first to really offer open claw hosting [20:47] and now Hermes hosting uh where it's [20:50] sort of all And And And other ones, [20:51] right? You can do N8N hosting. You can [20:53] do sort of one-click deploy, right? They [20:55] have Docker uh Docker installs for a lot [20:58] of different apps. That's not just And I [21:00] don't know what they what their start [21:02] was if they were strictly WordPress to [21:04] the to begin with or or not. Uh but [21:06] they've they've sort of expanded their [21:08] product line to keep up with demand, and [21:10] I wouldn't be surprised if [21:12] most of their [21:14] recent income is from stuff like [21:16] OpenClaw and Hermes. [21:19] >> Yeah. Have you, speaking of these [21:22] these agents, [21:24] um [21:25] have you gone down the route of OpenClaw [21:28] and Hermes? Have you any war stories [21:31] from trying to set one of those bad boy [21:33] bad boys up or do you ever use them in [21:35] your in your day? [21:37] >> Yeah, big Hermes fan. Yeah, I've been [21:39] I've been using that a lot. I have [21:42] several running. I do I have it running [21:43] on a VPS for our company [21:45] uh because, you know, we're a [21:47] distributed team, and so we have we have [21:50] it running on a VPS locked behind [21:52] Tailscale, and it's it's been great. It [21:56] lives in our Slack, and [21:58] it [21:59] updates us every day. It ingests our [22:01] meeting notes from Granola, and it's [22:04] it's like one of our employees almost, [22:06] and it's it's been very useful. Um it [22:09] updates our Notion and [22:11] updates our task lists, and [22:14] I mean, it it's very useful, I got to [22:16] say. And then I run it myself also [22:19] um [22:20] on my personal computer. [22:22] Talked with it on Discord and Slack and [22:24] Telegram, and I can, you know, chat with [22:26] it on my phone and stuff like that, and [22:29] it's getting easier and easier to set up [22:31] now. Now the the team behind it, News [22:33] Research, just launched the cloud [22:35] version where you can just set it up [22:37] also. I think it's like under 10 bucks a [22:39] month to host it. [22:40] Uh if you want to do that, and it's I [22:43] got to say, it's very easy to set up [22:44] now. The the the barrier to setting that [22:46] stuff up is pretty low now. Uh so you're [22:48] not dealing with a ton of environment [22:50] variables and API keys and stuff like [22:52] that. It's [22:53] it's getting easier and easier to set up [22:55] for non-technical people. [22:57] >> Yeah. You know, I haven't [22:59] I know I [23:01] definitely sound like, you know, old man [23:03] yells at the sky. I just haven't found [23:05] an absolute use case where I need to [23:09] spend the thinking power to set it up. I [23:12] set up OpenClaw once a few months ago [23:16] and I was just like, what am I going to [23:18] do with this thing now that I'm not [23:19] already achieving [23:21] with my other systems in place? [23:24] So, I like the WP Minute, I have what I [23:26] what what I call the JeffBot, right? So, [23:28] the JeffBot just summarizes content [23:33] articles from [23:35] other WordPress sites from a collection [23:37] of like, I don't know, 40 or 50 [23:39] RSS links that I have, right? And it [23:41] pings Slack, it summarizes it, pings [23:43] Slack with the link in the JeffBot [23:45] channel. [23:46] And it's all saved, that's all just like [23:48] Cloudflare running cron task [23:51] saving to a database so that [23:54] at any point I can [23:57] do more with it, like give me a weekly [24:00] updater, what were the most talked about [24:02] topics for this year or this quarter? [24:06] And to [clears throat] me, I'm just [24:06] like, yeah, it just feels cleaner to run [24:09] it that way versus [24:11] constantly thinking about managing [24:13] some kind of like agent infrastructure. [24:16] But I assume like as the stuff gets [24:18] easier and better, I'm also of the [24:20] mindset like Claude or Anthropic and [24:23] OpenAI will [24:24] come up with a native agent and I'm when [24:27] they do that, I'm just going to be like [24:29] freaking you I'll just pay you for it. [24:31] Like, I don't [laughter] want to I don't [24:32] want to go set this up on my own and run [24:34] another VPS infrastructure. You just run [24:36] it for me. [24:37] Um that's my mindset, but [24:40] um it's good to hear that you have some [24:41] success with that stuff. [24:43] >> Yeah, I mean, I it's also open source. I [24:45] think that's that's one of the things [24:47] that I mean, I I'm not going to I'm not [24:49] going to get I'm not going to pay [24:51] Anthropic API costs. Like, that that [24:53] that's insane. [24:55] >> Yes. [laughter] [24:55] >> They they they're and they're going to [24:57] want you to. They they they came out [24:58] with uh [25:00] I don't know what it was what they [25:01] called it like Claude can't keep up with [25:03] all their product releases but they they [25:05] they released some kind of thing where [25:06] you put it in slack and it responds to [25:08] things like that and it was it was like [25:10] a half-baked Hermes. [25:11] Yeah yeah [25:13] and and they got a they got a lot of [25:14] criticism for it cuz it was just [25:16] it was half-baked and it was just not [25:18] good. [25:19] And and Hermes is I don't know maybe I'm [25:22] a bit of a fanboy but I mean it's you [25:24] can switch models you can combine models [25:26] together on it. I mean it's [25:28] it's pretty amazing what you can do. I [25:30] mean just as like a test case I like was [25:32] able to just [25:34] just automate just go go in and do with [25:37] a new business I have set up a new [25:38] profile get everything up and running in [25:40] under an hour. [25:42] I mean you can't beat that. [25:44] >> What's the advantage like when you think [25:47] of it as [25:49] um [25:51] I'll frame how like if I I'm mostly in [25:53] Claude code [25:55] and then I use Codex [25:57] when I either hit a limit or I just want [26:00] to run [26:01] like documentation tasks [26:04] front-end edit tasks on some of these [26:06] projects like I'll use Codex for that [26:09] and a lot of like the the heavy work [26:10] I'll use Claude and then if I ever want [26:12] to [26:14] put something on [26:16] on a cron job that you that I want [26:18] analyzed every you know every day every [26:20] hour every week whatever that's when I'm [26:22] just like okay I got Wrangler [26:24] installed obviously on my machine go and [26:27] set up a cloudflare worker run this cron [26:30] x amount of times do the thinking shoot [26:33] it over to this either R2 or database [26:36] layer like I have my pockets of things [26:38] that I turn to and then when I want to [26:40] ask questions generally it's in chat GPT [26:42] when I'm kind of like brainstorming [26:44] stuff. Where does it the Hermes agent [26:46] fit in for you like how do you think of [26:48] it? Is it coding for you or is it just [26:51] like contest context tasks with memory [26:56] that you interface with it? [27:00] >> Yeah, it's I mean it's probably less on [27:02] the coding side. Yeah, it's more like [27:04] tasks and the and the cron stuff for [27:06] sure. Um I mean you can you can use it [27:08] to to code for sure, but I mean I don't [27:10] I don't hook up the really expensive [27:12] models to it. Like I'm running the [27:14] cheapo models through open router mostly [27:17] with it and [27:18] I mean have a couple of fallback models, [27:20] but I mean I have I think I have codex [27:22] hooked up to it as well, but um [27:25] mostly it's it's yeah, like you said, [27:27] it's like those cron jobs that need [27:29] automating and things like that. [27:32] Um so it's less less on the coding side [27:35] and more on like repetitive tasks and [27:37] you know, I I use something like [27:39] there's this really awesome service [27:40] called Compozio that lets you basically [27:43] connect all your different services [27:45] without needing to [27:47] get actual API keys. Like it's just a [27:48] one-click install and then Hermes has [27:51] access to your notion, to your slack, to [27:52] your whatever you want to give it access [27:54] to and you can fine-grain tune what it [27:56] has access to. So [27:58] basically just [28:00] one-click install what it has access to [28:02] and you're off to the races. [28:04] >> Mhm. That's interesting. [28:06] I have um [28:09] yeah, like a lot of the [28:11] uh we can talk about if you want to talk [28:13] about some of like the cool projects [28:15] maybe that you've built that that you [28:16] can talk about um cuz I know some cool [28:18] stuff that you can't talk about. [28:20] >> [laughter] [28:21] >> Uh but if you have cool stuff you do [28:22] want to talk about we can talk about it, [28:24] but [28:25] I I envision using I just haven't got [28:27] there yet. I I think and I think for an [28:29] agency owner, I've seen um my friend uh [28:34] Troy Dean from Agency Mavericks. [28:38] He's been all in on this [28:41] I guess this app called Victor v i k t o [28:44] r. I don't know if you've seen it. I [28:46] guess it's big in the agency space and [28:48] it's like uh [28:49] I mean, the way that I perceive it, the [28:51] the content that he puts out, it's like [28:53] uh a Hermes but for agency owners. [28:56] I mean, again, you'd have to go do your [28:58] own research on on that. [29:00] And I've seen him like loving it and he [29:02] uses it for like [29:04] client proposals and managing clients [29:06] and I see a world where like a Hermes or [29:09] an agent can and maybe should do that. [29:11] Like, here's all the knowledge about my [29:14] customer. Here's their website. Here's [29:16] their analytics. [29:19] You know, talk to me about what's [29:21] happening, you know, with their [29:23] with their site, with their product. Uh [29:25] follow them on social media and track [29:30] the social posts that they're putting [29:31] out so that I know [29:34] they've been happy with their like [29:35] they've been pointing to their site. [29:36] They've been telling people to go, you [29:37] know, sign up to the newsletter. Like, I [29:39] can see that as like [29:41] a massive value add for an agency owner. [29:44] I just don't know if it's [29:46] if it's there yet and how [29:48] accurate these agents are at this stage [29:51] in the game to like really replace what [29:53] I might call a [29:55] project manager or a customer success [29:57] person [29:58] in a typical company. [30:02] That makes sense. [30:03] >> Yeah. No, I mean, I [30:04] I mean, I'm never in the business of uh [30:06] replacing people. [30:08] >> And I was going to say that as well. [30:10] >> [laughter] [30:10] >> Yeah. No, I mean, and and I've had that [30:13] conversation with with other business [30:15] owners where they're like, "Oh, can I [30:16] just use this Hermes agent to replace my [30:19] project manager?" I'm like, "Well, [30:20] that's not really That's not really what [30:22] I'm trying to propose here." [30:24] >> Yeah, yeah. [30:24] >> Uh it's more like augmenting or trying [30:26] to, you know, make sure they're not [30:28] making mistakes and things like that. [30:30] Uh I mean, I think there's a lot of a [30:32] lot of the the work that project [30:34] managers do where it's [30:36] uh taking taking meeting notes and [30:38] turning it into uh tasks for your team. [30:41] I think that that kind of stuff is is [30:43] definitely there. Um [30:45] and you know, making sure that there's [30:47] not context drift, um you know, that [30:50] that you're not uh you know, having [30:53] having stuff threads get lost and then [30:56] you know, dates get changed and and [30:58] there's not, you know, stuff that's, you [31:00] know, getting lost in in the in the [31:02] the hustle and bustle of agency life [31:04] that we all know about. Um that I think [31:07] that's an important thing that that [31:09] these agents can really do and that's [31:11] something that I've seen in in my [31:12] day-to-day life with uh with this Hermes [31:15] agent that we have in in our company. [31:17] It's really helped with that kind of [31:18] stuff, for sure. Um I think that [31:22] that you're I mean, you're right to not [31:25] really put your full trust in these [31:26] things yet. I think that's that's an [31:28] important thing, you know, [31:29] making sure that there's human in the [31:31] loop thing that you're that you're still [31:33] like, you know, human is approving what [31:35] these agents are doing. That's that's [31:36] good. Um there was a there was a good [31:39] article that the Shopify team put out [31:41] about their agent called River uh a [31:44] couple months ago that that lives in [31:46] their Slack and it's all about how [31:48] they've also just moved everything out [31:50] of Slack DMs into public channels. And [31:53] that's really helped the agent just like [31:56] become more aware and get better context [31:59] because if everything lives out in [32:00] public channels and less in DMs, then it [32:03] just it gets smarter and smarter and has [32:05] better context about how your company [32:07] works and that's I recommend people read [32:11] that article and and that's that's sort [32:13] of how I've [32:14] you know, kind of built my Hermes agents [32:17] moving forward for for my clients as [32:19] well. Um because that's I think these [32:22] things are only as good as the context [32:24] you give them and the tools you give [32:25] them, right? [32:27] So. [32:28] >> Yeah. I mean, running a [32:30] you know, my own social experiment which [32:32] which you know about with like a the [32:33] civic uh of tooling that I've been uh [32:36] working on for a while, right? And it's [32:39] grown beyond I I haven't really updated [32:41] you on it, but um [32:44] you know, I had like that 311 [32:46] non-emergency reporting tool, potholes, [32:49] etc., etc. That's been running for a [32:51] while, but over the last month or so, um [32:55] I've had that meetings app, and the [32:58] meetings app just ingests the uh the um [33:02] transcripts from YouTube and gives [33:05] [clears throat] place gives people a [33:06] place to uh search through all these [33:08] meetings through the transcripts, and I [33:10] do the timestamp topics and all this [33:11] stuff, and it sort of sheds light on all [33:13] like these [33:14] boring, long government meetings that no [33:16] one's ever going to watch, and then like [33:18] hopefully distilling it down to, you [33:20] know, simple chapters and keyword search [33:22] for people to find. [33:24] But, over the last um [33:25] month or so, I've [33:27] built out a pretty aggressive pipeline, [33:30] and now I've been experimenting [33:33] um with taking all of that context and [33:37] content and having AI sit in the middle. [33:40] Um I'm using [33:43] Oh god, what model am I using for the [33:45] for the [33:46] blog posts? [33:48] I think I am using Claude for the blog [33:50] posts. Uh I have to go I don't even [33:52] remember off the top of my head, but [33:54] anyway, I've put WordPress in the [33:55] middle. [33:56] So, my [33:58] my [33:59] or not in the middle on the on the on [34:00] the front end. Um and what it'll do is [34:04] through cron and as these all these [34:06] different apps fire, [34:08] um [34:09] it will take that information and then [34:11] post it to WordPress. [34:12] And that's giving [34:15] it's allowing me to just to experiment, [34:17] but it's allowing me to say, "Hey look, [34:18] if I want to build a front end for all [34:20] of this content, [34:21] and eventually get [34:23] a human in the loop, WordPress is [34:25] actually a great platform for that. [34:27] Because now I can generate all of this [34:29] content through AI, post it on [34:31] WordPress, [34:33] and because I simply don't trust all of [34:35] AI yet, and there are inaccuracies [34:37] happening, how can I give citizens the [34:41] ability to go in and like Wikipedia [34:43] style say, "All right, no, that's that's [34:45] the wrong name or that's the wrong [34:47] street." Or give people editorial access [34:50] to kind of like own this AI generated [34:52] news? [34:53] Because it is coming, and it's going to [34:54] come at like probably like the national [34:56] global level first before it hits like [34:58] the local markets. [35:00] And And this has been a very interesting [35:02] um experiment because it's it's [35:05] interesting to see like what AI chooses [35:07] as hot topics. And And it's doing [35:10] actually a pretty good job, but [35:13] clearly getting a lot of like the names [35:15] and like local vocabulary wrong. Um [35:18] because I've been relying strictly on uh [35:21] the YouTube transcript. [35:24] And [35:25] that's been really cool to see. I have [35:27] it generate an automatic podcast, and it [35:29] is interesting to see [35:31] every week listening to my my week ahead [35:34] robot, uh thanks to 11 Labs, tell me [35:37] what's about to happen. [35:39] And then at the every Friday a weekly [35:41] recap, and listening to a robot go, [35:44] "Here's what happened throughout the [35:45] city on Friday." And I have a friend at [35:49] the local paper, he's the editor for the [35:51] local paper, and um [35:54] he's kind of keeping an eye on it, and [35:55] I'm like, "What do you think about this, [35:56] Dan?" And he's just like, "You're [35:58] literally building a thing to replace my [36:00] job." And I'm like, "I'm not trying to. [36:02] Like, I'm trying to build a system that [36:04] is going to [36:07] take all this this This is the long way [36:08] of getting to like why these agents are [36:10] important. There's so much so much [36:12] content that's happening in a business, [36:16] in a local government, that you just [36:18] don't have the manpower or the time to [36:21] watch all of it. [36:23] And this is the major unlock for me [36:26] anyway that AI [36:27] can surface for you. Like at an agency, [36:30] you might have a meeting for your [36:31] engineering team that's going to impact [36:33] your client, but your designers don't [36:34] know about it, your project managers [36:36] don't know about it. And if no one's [36:37] paying attention, maybe AI can [36:40] autonomously say, "Oh, you know, Brendan [36:42] said he's going to launch a new plugin [36:43] for this client. The sales team said [36:45] this client's leaving. [36:47] Maybe we should warn everybody that says [36:50] there's a disconnect here." And, you [36:52] know, no different than in local [36:54] government, you know, one meeting says [36:55] this, another meeting says that. No [36:57] one's paying attention, so nobody knows. [36:59] So, anyway, long way of getting to [37:02] I definitely see an advantage for these [37:04] agents. [37:07] >> Yeah. Well, no, that's [37:09] I didn't even think about that. That's a [37:10] very good point. You've You've You've [37:12] told a very compelling story there. I [37:14] might have to steal that. Um [37:16] That was very very well put there, Matt. [37:19] Thank you. [37:20] >> Yeah. [37:20] >> Uh that's No, that's exactly That's [37:22] exactly what we have uh sort of in our [37:24] pipeline where all of our meetings get [37:26] transcribed. We put the original [37:27] transcript into GitHub. And then we put [37:30] what project it's attached to, what was [37:33] said, who was there in the meeting, and, [37:35] you know, basically what was said in the [37:36] in the meeting, and what what projects [37:38] and tasks touched that meeting, right? [37:41] And then, [37:42] you know, it's it's basically updated [37:44] every single day at 4:00 p.m. [37:46] Uh and, [37:48] you know, it's it's our second brain for [37:51] our company, right? And that's that's [37:52] updated every single day, and it's it's [37:55] been crucial moving forward for [37:58] you know, making sure that our Notion [38:00] stays up to date, and making sure [38:01] everyone is on the same page, because [38:03] You're exactly right. We Not everyone is [38:05] in in every meeting. [38:07] >> Yeah. How do you feed it I mean, if you [38:10] can say, like how do you feed it [38:13] information to learn from? Is it all [38:15] just like generated by Slack? Is that [38:17] the catalyst for your Hermes agent to [38:20] know everything that's happening, or are [38:22] you feeding like from all [38:24] angles? [38:25] >> All angles, yeah. So, it gets it gets [38:27] Slack context uh through the Slack API, [38:30] and then it gets uh Granola meeting [38:33] notes, which is uh a meeting note that [38:35] you So, every meeting that we're in, [38:37] someone's running Granola. [38:39] And so, it gets that API access. [38:42] Uh and then it has Notion access through [38:44] that API. So, that trifecta is basically [38:48] all the company information. [38:51] >> Yeah. [38:52] Do you give it like oh, new contract new [38:54] customer contracts? So, like does that [38:55] is that how it is customer aware or you [38:57] not at that level yet? [39:00] >> Yeah, I mean customer customer [39:01] information No, so new projects that [39:03] come in are are put into Notion. So, [39:05] that's what we use for projects. That's [39:07] basically the CRM. [39:09] Uh you know, my uh [39:11] the agency owner wants to to give it [39:13] access to his email, but there's a lot [39:15] of confidential stuff in there. And I'm [39:17] like, well, okay, let's hold off on [39:18] that. [39:19] Or let's figure out some some way to not [39:22] give it everything in your email. And [39:24] then And yeah, so that's that's the next [39:26] frontier. [39:27] >> Yeah. [39:28] Yeah, that's stuff that I'm not [39:30] uh [39:31] you know, willing to give up. And [39:33] you [clears throat] know, another [39:34] advantage of these agents, which I've [39:36] heard a lot about and I know is a lot of [39:37] promise for agency owners, is like, boy, [39:39] wouldn't it be great if it would just go [39:41] do like every when anything ever sounds [39:43] like too good to be true, then it's too [39:45] good to be true. But I think like a lot [39:46] of people say, "Oh, I've got these [39:47] agents that just like cold email [39:50] everybody or cold contact people." [39:53] And of course, like as a sales person, [39:55] in my head, I'm like, "That's obviously [39:57] I want that. Give it to me." And you [39:59] know, I just want to start doing that. [40:02] But I just saw a report today, I think [40:04] it was like 40% of like LinkedIn posts [40:06] or they just they like a survey or some [40:09] outlet did a survey where 40% of AI or [40:12] LinkedIn posts are just AI, you know, [40:14] slop that these people are just [40:15] unleashing on on LinkedIn. [40:18] And it's just like, man, when [40:20] like is that the is that the future of [40:22] communication? Just more spam and more [40:26] bots, you know, hitting your inbox? [40:29] Or is there a [40:31] a way that like the future of [40:33] communication is [40:34] contact my agent. [40:36] If if I tell my agent I'm looking for [40:39] something [40:40] or my agent knows I have a particular [40:41] pain point, [40:43] then the agent lets it through to me. [40:46] And just have these agents talk to each [40:47] other. Stop hitting my inbox. Just talk [40:50] to my agent. And if my agent says, "Oh, [40:52] yeah, Matt's got that problem, lead [40:54] generation that he's hoping for you to [40:56] solve." [40:57] Boom, land on his land on his desk. Um I [41:01] don't know. That's that's the hopeful [41:02] world that I'm that I'm looking for, but [41:04] I think a lot of folks unleashing [41:06] unleashing these agents [41:07] um [41:09] are just like, you know, taking an easy [41:11] route. This is not going to pan out for [41:12] them. [41:14] >> Yeah. I mean, I certainly know my inbox [41:17] is just full and full of spam. And [41:19] LinkedIn is [41:20] not only are the posts a bunch of spam, [41:23] but the comments are just AI-generated [41:25] as well. I mean, it's [41:26] >> Yeah. [41:28] >> I don't know if there's an end in sight [41:30] to it. And I I I can't I just can't tell [41:32] what's [41:33] I People assure me that the Oh, this is [41:36] a human post. And it's like, I don't [41:38] believe you. So [41:40] >> [laughter] [41:42] [gasps] [41:42] >> Uh [41:43] man. Yeah, the uh [41:46] Yeah, I don't know. I mean, that's a [41:48] whole different conversation. It's like [41:49] trust in content and stuff like that. [41:51] And one of the things that I'm trying to [41:53] do with my social experiment is to make [41:55] it uh the next thing I'm going to be [41:57] going to do is have a an AI rating. Like [42:00] there's disclaimers everywhere on the [42:01] site that this is AI-generated content. [42:04] And then if you feel the need to tell me [42:06] how well it did, you know, you can hit [42:08] this, you know, ranker. [42:11] You know, one out of five. You know, [42:12] between one to five, like where how did [42:13] how did this AI bot do [42:16] um to try to like improve that and like [42:18] feed that back to uh the machine, but um [42:23] yeah, it's a whole new social [42:24] experiment. Let's just round out the [42:25] episode uh talking about WordPress. Have [42:27] you Is there anything that you're [42:29] looking ahead to in WordPress world [42:31] going [42:33] they need to do this in order to win? Is [42:36] there something that you're on the [42:37] sidelines going [42:39] it's getting a little bit better in [42:40] terms of like the AI adoption in [42:42] WordPress, but if they do this [42:45] like this is what I need as somebody who [42:48] lives in the, you know, WordPress space. [42:50] Is there something that's got you on the [42:52] hook? [42:54] >> Uh [42:55] >> [sighs] [42:55] >> I mean, that's a good question. I think [42:56] there you know, there's been there's [42:58] been quite a lot of improvements, I [43:00] guess. I I guess 7.0's new since we've [43:03] since we've last spoke. Uh you know, [43:05] there's there's quite a lot in there to [43:07] unpack. [43:08] Uh I I don't know. I was I was trying to [43:11] just connect up the [43:13] the connect the generic MCP the other [43:15] day and to took me much longer than I [43:18] wished. [43:19] Um and [43:21] and I don't know if that was just [43:22] because I don't do it enough anymore, uh [43:26] but I I wish that that was easier to [43:28] expose the abilities that are available [43:32] and to get [43:33] to you know, just to to get the [43:34] abilities API really just [43:38] have that easily available to everyone [43:41] because that's super powerful stuff, [43:42] right? I mean, uh you know, I don't [43:45] know, like a Meta Box just had a big [43:47] release where they have [43:49] the abilities API. I know that the you [43:51] know, I'm sure Gravity Forms has the [43:52] abilities API exposed and other form [43:55] builders have abilities API. The [43:57] abilities API is great. And I think you [44:00] still need to have the MCP plugin to to [44:04] use those. [44:05] And [44:07] and I you might need to do other steps [44:09] to do that, but I just think it's too [44:12] too cumbersome right now, perhaps. I [44:14] would I would like to see that become [44:17] easier [44:18] to work with um [44:20] and to yeah, just in general. And I [44:23] think that would that would help a lot [44:25] out in terms of getting agents uh it [44:27] working with your your WordPress site [44:29] easily. So. [44:31] >> Yeah, I'm 100% [44:32] >> If we're talking about the AI aspect of [44:34] WordPress. [44:35] >> Yeah, I mean [44:36] yeah, right I'm right there with you. [44:37] That's that's 100% mine. Like that's [44:39] going to be the thing that is figured [44:40] out next to um you know, demystify that [44:44] MCP setup. We have uh abilities in [44:46] Gravity Forms. We're just [44:48] waiting for it's not publicly available [44:50] right now, but it'll be by the end of [44:52] the month. We're been just waiting for [44:53] this next version of ours to to go out, [44:56] but yeah, it's it's amazing. And I I see [44:59] a lot of the the haters um [45:02] especially like the static site builders [45:04] talking about like, "Oh, look how fast [45:06] I can build things with, you know, XYZ [45:10] uh tool and WordPress can't do this." [45:13] And [45:14] I sort of chuckle every time I see that [45:15] cuz I'm like, "Well, [45:17] as soon as WordPress does this, [45:20] like I think you [45:21] as soon as WordPress unlocks like ease [45:23] of use of setting up MCP or it's shipped [45:25] in core [45:26] uh or even if it's just like a more um [45:29] uh what do you call it? Recommended [45:31] plugin where you just click and install [45:33] the MCP service or they make it part of [45:35] the abilities or the API page right now. [45:39] Once they just like figure that part out [45:41] and then it allows all the developers to [45:43] have their abilities like Gravity Forms [45:45] or your favorite plugin or whatever. [45:47] Like we're going to be able to [45:50] talk in English to the WordPress site [45:52] and it'll do what we want it to do. [45:54] So, like all of this stuff where [45:55] everyone's like, "Oh, Astro does all [45:57] this. Like you can't Yeah, it's so easy [45:59] for me to just type in Codex and build [46:00] this thing." WordPress is going to do [46:02] that. Like, on one side I'm like, when [46:05] the hell is WordPress going to do this? [46:07] But as soon as it does, like, we're all [46:09] going to be living in a world where you [46:11] are just talking and building a [46:12] WordPress site. I don't know why [46:14] everyone's like jumping ship because of [46:17] it or because it's not here yet. It's [46:19] going to get here. It's just maybe [46:20] people can't wait or they just want a [46:22] new tool. [46:23] But it's going to be a different world, [46:24] man, when you can just say [46:27] create this site, put in this dummy [46:29] content, install gravity forms, put in [46:32] dummy entry gravity form stuff, create [46:34] this form for me, install WooCommerce, [46:37] yada yada yada. [46:38] And [46:39] um [46:40] yeah, it's like it's just we're like it [46:42] feels like we're right there. [46:44] Um but we're just haven't pushed forward [46:46] yet. Um but I'm I'm excited for it when [46:49] we do have it. [46:50] >> And we can do that [46:52] in command line or you know, within a [46:54] terminal locally very easily, I think [46:57] now. [46:58] Um [46:59] and I think getting that [47:01] safely done on a VPS cuz I think that's [47:04] probably how the majority of WordPress [47:06] users are interacting with their sites. [47:09] Uh so I think that would be you know, [47:10] the next unlock. I don't think most [47:12] people are going to be doing that via [47:13] SSH. [47:15] Uh [47:15] so that would be the next unlock maybe [47:17] after that. [47:19] Um so and and making sure that it's not [47:21] just [47:22] in the hands of hosting companies would [47:25] probably be [47:27] good because I think that you know, a [47:28] lot of hosting companies probably are [47:30] wanting to monetize that and keep that [47:33] on you know, which is fine. They can do [47:35] that, you know, fine. They can do that. [47:37] I don't care. But you know, I think it'd [47:38] be good to to have that as some some [47:41] feature in core because I think it's [47:43] important for WordPress to [47:45] to stay to stay relevant in that front [47:47] to keep it [47:49] somewhat top of mind for people and and [47:51] easier for your average person. [47:54] >> Yeah, I just use the app. [47:57] Um [47:59] What's it called? Like the app password [48:01] do do do do. Of course I'm going to [48:02] forget the name of it. You can set it up [48:04] inside your um [48:06] >> [clears throat] [48:06] >> inside your user account. And I run that [48:09] as uh a local ENV file or in my in local [48:13] environment file and I can just talk to [48:17] Cloud that way without even like the [48:18] traditional MCP setup. And it'll just [48:21] access the content and like when I need [48:23] to make big changes of content or create [48:26] a page or a set of pages [48:28] uh or update a privacy page like it'll [48:30] just go and do it. And um it's just a [48:33] one-to-one relationship, but yeah, it's [48:36] it's so it's great. I don't know what [48:38] the hell everyone's complaining about. [48:40] >> [laughter] [48:41] >> Yeah. [48:41] >> I know what everyone's complaining [48:42] about. Uh Brandon, we're going to wrap [48:44] it up [48:45] today. [48:45] >> [laughter and gasps] [48:46] >> We're going to wrap it up today. Uh [48:47] thanks for hanging out on the Agency [48:49] Action podcast. You've you've hit every [48:52] every show in the the WP Minute uh [48:56] stable. You're across all shows. You [48:58] hold that [48:59] Congratulations. [49:01] Um you're I'll put your face above Mark [49:04] Semansky. [49:06] >> There we go. That's That's going to [49:07] count for something. [49:09] >> It's going to count for something. [49:10] Uh Brandon, where do you want folks to [49:11] go to say thanks? [49:14] >> You know, you you can find me on [49:15] YouTube. You can find me on various [49:17] socials. Brandon O'Connell. [49:19] Search for that. [49:21] / I don't know. Whatever. Uh -WP I don't [49:24] know. [49:25] >> [laughter] [49:27] >> I I'm not I'm not the Brandon O'Connell [49:28] artist. That's that's all I know. [49:30] >> Yeah. [49:31] >> I'm the Brandon O'Connell developer [49:33] WordPress extraordinaire, whatever it [49:36] is. [49:37] Uh [49:38] on the various uh platforms. So. [49:40] >> You've always got some uh really [49:43] in-depth video that that we all look [49:45] forward to. Uh thanks for sharing all [49:47] the stuff that you do. Everybody else uh [49:50] Agency Action podcast part of the WP [49:52] Minute Network, where um every month we [49:55] have we've been having member meetups [49:56] for the last three-ish months. You can [49:58] join as a free member by taking one of [50:00] our free courses. Go to the [50:02] wpminute.com. [50:04] Uh click on the courses button, pick one [50:05] of our free courses, join for free. [50:07] There's a Slack button, you can join the [50:09] Slack. We want you in there. And then [50:11] you can meet up with us every month to [50:12] talk about uh cool agency stuff. There's [50:14] always a new topic. Uh editor-in-chief [50:17] uh Eric Karkovack runs that stuff. So, [50:20] join us for free in the membership in [50:22] Slack. We'll see you in the next [50:24] episode.  ## Highlights - if you have just like utility products um those will probably the be the first to go... if I need something to replicate on all my client sites with all the add-ons and connections. Do I want to risk that for 200 bucks a year to buy Gravity Forms? Probably not. - Note: 📚 Utility plugins face more AI substitution risk than ecosystem plugins · - Mid-episode, during the WordPress-product-risk discussion - Directly actionable for agency owners deciding which client plugins to defend vs. treat as replaceable · #atom/literature/h1 - we have these open weight model like GLM... very capable. And so... they're seeing these open weight models come from behind thinking that that's changing their their calculus - Note: 📚 Open-weight model competitiveness is reshaping frontier-lab pricing tiers · - Discussion of Anthropic's Opus/Mythos/Fable tiering strategy - Supports hypothesis 5 above; testable via public pricing/trial-duration tracking · #atom/literature/h2 - Hostinger... They've kind of embraced... the AI agent craze... they were one of the first to really offer open claw hosting and now Hermes hosting - Note: 📚 Hosting companies are pivoting from web hosting to agent/inference hosting · - Follows the "hosting companies as malls" analogy - Relevant to agency owners choosing hosting partners for client agent deployments · #atom/literature/h3 - Buy-vs-build economics inverted, not eliminated, by cheap AI coding - Note: Holds when: the existing product has deep integration/add-on surface area (Gravity Forms example) that a custom build would need to replicate. · - [[Hosting-as-inference-infrastructure]] — because contrasts on which layer of the stack captures value as building gets cheaper · #atom/zettel/z1 - Find and read Shopify's public engineering write-up on their internal agent "River" (public-channel-context strategy) — referenced verbally, URL not given in source #xref - Note: ✏️ Fleeting capture #atom/fleeting/f1 - Look into Troy Dean / "Viktor" as an agency-specific agent tool — mentioned as popular in agency circles but not independently verified in this episode #xref - Note: ✏️ Fleeting capture #atom/fleeting/f2 - ** Matt's claim that "40% of LinkedIn posts... are just AI slop" is cited as "a survey... some outlet did" without naming the source or methodology — presented with more confidence than the sourcing supports. Fix: name the source and sample size before using it as evidence for a broader claim about communication's future. - Note: 🔍 Fallacy: Hasty Generalization #fallacy/hasty-generalization