# Automate, Build, or Buy? The New Agency Playbook **Source**: [[3. Resources/Readwise/Articles/Automate, Build, or Buy The New Agency Playbook]] > [!question] Lingering Questions > - What specific friction points make WordPress MCP setup "cumbersome" today — protocol-level or plugin-implementation-level? Brendan names the pain but not the root cause. > - Is there quantitative data behind the cited "~40% of LinkedIn posts are AI slop" claim, or is it an unsourced anecdote functioning as evidence? > - How is Hermes's context-ingestion (Slack + Granola + Notion) governed with respect to sensitive data, given email access was explicitly withheld? > [!tip] Article Ideas > - A decision matrix formalizing the episode's recurring implicit logic: utility plugin vs. ecosystem plugin, local vs. hosted inference, augment vs. replace project manager — three separate binary framings that keep resurfacing across the conversation. > [!link] Cross Reference > - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — buy-vs-build economics under cheap AI coding is a direct instance of this MOC's core question > - [[3. Resources/Permanent Notes/Local Inference Is a Procurement Spectrum Not a Binary]] — this episode's own RTX 4090/Gemma testing ("dog slow" vs. hosted) is fresh evidence for that existing claim > [!quote] Vocabulary > - **Hermes** — an open-source, always-on personal/team agent Brendan self-hosts on a VPS, ingesting Slack, Granola meeting notes, and Notion to push daily updates > - **Composio** — the connector layer letting Hermes reach Slack/Notion/etc. without manual API key setup, lowering the barrier to non-technical agent setup ## Bib capsule Matt (guest-hosting) and returning guest Brendan O'Connell discuss how WordPress agency AI workflows have shifted over roughly six months: automation tooling (n8n) blending with, not being replaced by, AI coding agents; always-on personal agents (Hermes) emerging for internal ops; and WordPress's MCP/Abilities API friction as the ecosystem's real remaining blocker. ## Author's thesis 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. ## Key arguments (≤3) 1. **Cheap AI coding inverted the buy-vs-build calculus without erasing product moats** — utility plugins face real AI-driven substitution risk; ecosystem-anchored products with deep integration surface (Gravity Forms) are more insulated. 2. **Compute, not model weights, is the real bottleneck** — local consumer-GPU inference tested "dog slow" against hosted models for both hosts, meaning whoever controls affordable, fast inference wins the next phase. 3. **Hosting companies are starting to pivot from web hosting to agent/inference hosting** — framed via a "malls that didn't reinvent themselves died" analogy, with Hostinger's OpenClaw/Hermes hosting cited as an early instance. ## Claim atoms ### 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." Directly actionable for agency owners deciding which client plugins to defend vs. treat as replaceable. ### 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" Discussion of Anthropic's Opus/Mythos/Fable tiering strategy — testable via public pricing/trial-duration tracking over time. ### 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" Follows the "hosting companies as malls" analogy — relevant to agency owners choosing hosting partners for client agent deployments. ## Permanent candidates - **Ecosystem lock-in insulates products from cheap AI substitution** — the strongest, most reusable candidate; promoted below. - The local-inference-lags-hosted-models observation isn't promoted as a new note — it's fresh evidence for the vault's existing [[3. Resources/Permanent Notes/Local Inference Is a Procurement Spectrum Not a Binary]] claim; connect rather than duplicate. ## Processing checklist - [ ] Find and read Shopify's public engineering write-up on their internal agent "River" (public-channel-context strategy), referenced verbally but not linked in the source - [ ] Look into Troy Dean / "Viktor" as an agency-specific agent tool, mentioned but not independently verified in the episode - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — buy-vs-build economics and agency-scale AI adoption