Mythos Is Here, It’s Time to Start Tokenmaxxing

Mythos Is Here, It’s Time to Start Tokenmaxxing ![rw-book-cover](https://img.youtube.com/vi/3sTu8sSLVfg/maxresdefault.jpg) 🔗 **Source:** [Mythos Is Here, It’s Time to Start Tokenmaxxing](https://www.youtube.com/watch?v=3sTu8sSLVfg) > [!note]+ Document Notes > tags: [ai-agents, token-optimization, claude-code, agent-orchestration, development-automation, cost-efficiency] > author: [MISSING] > source: YouTube > reference: https://www.youtube.com/watch?v=3sTu8sSLVfg > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: Mythos is here, it's time to start tokenmaxxing >  ## 🧭 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 - Fable (Mythos with safeguards) is temporarily available on Claude Code pro/max tiers until June 22nd - The subsidized plans offer up to $8,000/month of inference for just $200, creating a massive arbitrage opportunity - Creator demonstrates burning through $4,358 of inference in 10 days using aggressive workflows and multi-account strategies - Key strategies include rate limit optimization, account switching, workflow automation, and remote machine orchestration - The temporary nature of this access creates urgency to maximize usage before the model moves to more expensive tiers ## 📖 Main Content The video focuses on maximizing usage of Claude's Fable model during its limited availability on subsidized subscription plans. The creator has burned through thousands of dollars worth of inference in just 10 days, demonstrating both the power and cost implications of advanced AI models. ### Rate Limit Optimization The system has two key limits: a 5-hour session limit and a weekly limit. The creator shows how to game these by triggering the countdown timers early through simple messages, ensuring maximum reset frequency. He even automates this with cron jobs to maintain active rate limit countdowns. ### Multi-Account Strategy By maintaining two Claude Code accounts on different reset schedules, the creator can seamlessly switch between them when hitting limits. The CLI tool makes this trivial - just run `/login` and authorize in a different browser profile. Ongoing workflows automatically route to whichever account token is active. ### Workflow Automation The video showcases expensive but powerful workflows that spawn dozens of sub-agents for tasks like PR review and code auditing. A single workflow can burn through 1.8+ million tokens by running 8+ agents in parallel, quickly consuming daily and weekly limits. ### Remote Machine Setup A Mac Mini running agents remotely allows work to continue when closing laptops. The creator accesses it via SSH, screen sharing, and T3 Code's remote system, enabling longer-running jobs and phone-based task initiation. ## 🗒️ Notes #### ⏱️ Rate Limit Gaming Strategy - Two separate limits: 5-hour session and weekly total - Timers only start counting down after first message - Simple "hi" message followed by immediate stop triggers countdown without real usage - Automated with cron jobs running every 5 hours using claude-p CLI #### ⏱️ Account Switching Mechanics - Two accounts on different reset schedules (Wednesday/Thursday) - CLI login updates all terminals on same machine instantly - Mid-workflow account switches work seamlessly - next tool call routes to new token - Prevents workflow failures when hitting limits (especially painful for 100+ sub-agent workflows) #### ⏱️ Workflow Token Consumption - Single workflow: 1.8M+ tokens in under 30 minutes - 8 parallel agents can hit 5% of weekly limit in minutes - Example: PR review workflow with 13 audit agents + judge agents + verification - Weekly limit appears to be ~4x the 5-hour limit based on testing #### ⏱️ Practical Applications - Daily PR triage across multiple repositories - Automated ranking of merge-ready PRs with justification - Cross-model review (Codex writes, Mythos reviews) - HTML plan output for readable agent summaries #### ⏱️ Remote Orchestration Setup - Mac Mini with Codex, Claude Code, and Hermes Agent - Access via SSH (with Tailscale for remote), screen sharing, T3 Code remote - Enables phone-based task initiation and laptop-free operation - Allows much longer exploratory agent runs #### ⏱️ Advanced Automation Patterns - Agents reviewing each other's work automatically - Browser automation with screen recording for verification - Self-triggered maintenance loops checking repos every 5 minutes - Integration with monitoring tools (Sentry, DataDog) for autonomous bug fixing ## 📚 References - [Claude Code](https://claude.ai) - [T3 Code](https://t3.gg) - [Render Cloud Platform](https://render.com) - [Tailscale VPN](https://tailscale.com) - [Raycast Notes](https://raycast.com) ## 🗣️ Paraphrase A developer demonstrates extreme optimization of Claude's Fable model access during its limited availability on subsidized plans. Through multi-account management, rate limit gaming, and automated workflows, he burns through thousands of dollars of inference value for $200 monthly subscriptions. The approach involves setting up remote machines for continuous operation, automating agent orchestration, and pushing the boundaries of what's possible with current AI tooling before the pricing model changes. ## 📰 Newsletter 🚀 **Token Maximization Alert!** Claude's Fable model is temporarily available on pro plans until June 22nd - and developers are going wild with the arbitrage opportunity. One creator burned $4,300+ of inference in 10 days for just $200, using multi-account strategies and automated workflows. Key takeaways for WordPress developers: Consider how agent orchestration could automate your plugin review process, code audits, and issue triage. The remote machine setup shown could be perfect for running continuous site monitoring and optimization agents. While this extreme usage isn't sustainable long-term, it showcases the potential of AI-assisted development workflows that could transform how we maintain WordPress projects. Time to experiment before the pricing changes! ## ✅ Action Items #### DevRel - Experiment with AI workflow automation for community PR review and triage - Document agent orchestration patterns for developer education content - Create demos showing multi-model collaboration (Codex writes, Claude reviews) - Develop guidelines for sustainable AI-assisted development practices #### WordPress - **Core**: Test automated PR ranking and merge-readiness assessment workflows - **Community**: Explore agent-assisted meetup planning and speaker matching - **Training**: Develop AI-assisted code review curriculum for contributors - **Documentation**: Experiment with automated documentation quality audits - **Support**: Test automated ticket triage and response prioritization #### Community - **Maintainers**: Share sustainable patterns for AI-assisted project maintenance - **Contributors**: Document onboarding automation using AI agents - **Sponsors**: Analyze cost-effectiveness of AI tooling investments vs. traditional development ## 🧪 Hypotheses - Multi-account rate limit optimization can increase effective AI model access by 200-400% during subsidized periods - Workflow automation with parallel sub-agents provides better code review quality than single-threaded approaches - Remote machine orchestration enables 10x longer AI agent run times compared to laptop-bound workflows - Agent-to-agent review cycles produce higher quality code than single-agent approaches - HTML plan output format improves human comprehension of complex agent workflows by 60%+ - Phone-based task initiation increases developer productivity by reducing context switching friction ### 📊 Methodology - A/B testing of single vs. multi-account usage patterns over 2-week periods - Token consumption analysis across different workflow patterns (single-threaded vs. parallel) - Quality metrics comparison between human-only, single-agent, and multi-agent code reviews - Time-to-completion studies for laptop-bound vs. remote machine agent workflows - User comprehension testing of markdown vs. HTML agent output formats ### ❓ Questions - How will pricing changes affect sustainable AI development workflows after the subsidized period? - What are the optimal sub-agent orchestration patterns for different types of development tasks? - How can smaller development teams replicate these patterns without multi-account complexity? - What quality control mechanisms prevent runaway agent behaviors in automated workflows? - How do different model combinations (GPT-4, Claude, Codex) perform in collaborative agent setups? ### 🔮 Future Research - Development of cost-optimization frameworks for post-subsidy AI development - Agent workflow pattern libraries for common development tasks - Quality metrics and validation systems for AI-generated code at scale - Integration patterns between AI agents and traditional DevOps tooling - Sustainable resource allocation models for AI-assisted development teams ### 💡 Implications - **Practice**: Development teams need new resource management strategies for AI tool usage - **Policy**: Organizations require governance frameworks for AI agent autonomy and spending - **Security**: Automated code generation at scale needs enhanced security review processes - **Education**: Developer training must include AI workflow optimization and orchestration - **Economics**: The AI development subsidy model is unsustainable, forcing evolution toward usage-based pricing ### 🧠 Learning Styles Assessment Reading level: College level (Flesch-Kincaid Grade 12-14) due to technical terminology and complex workflow descriptions. #### 👀 Visual learners - Present: Terminal screenshots, workflow diagrams, account switching demonstrations - Improve: Add flowcharts of agent orchestration patterns, visual rate limit dashboards, architecture diagrams showing remote machine setup #### 🔊 Aural learners - Present: Video format with narration explaining complex workflows - Improve: Add audio summaries of key optimization strategies, podcast-style discussion of implications #### 🗣️ Verbal learners - Present: Clear definitions of rate limits, tokens, workflows; step-by-step explanations - Improve: Add glossary of AI development terms, comparison tables between different approaches #### 🧩 Kinesthetic learners - Present: Hands-on CLI commands, copy-pasteable automation scripts - Improve: Add interactive tutorials, step-by-step setup checklists, configuration templates #### 🧠 Logical learners - Present: Token consumption metrics, cost-benefit analyses, systematic optimization strategies - Improve: Add decision trees for account switching, ROI calculations, efficiency benchmarks #### 🧑‍🤝‍🧑 Social learners - Present: Community sharing of optimization strategies, collaboration between multiple AI agents - Improve: Add discussion forums for optimization techniques, peer review of automation setups #### 🧘‍♀️ Solitary learners - Present: Self-directed experimentation approach, individual account optimization - Improve: Add reflection prompts on personal development workflow improvements #### 🌿 Naturalistic learners - Present: Resource optimization patterns, sustainable vs. unsustainable usage models - Improve: Add analogies to natural resource management, ecosystem thinking about AI development ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive timestamps - Improve: Add alt text for terminal screenshots, ARIA labels for workflow diagrams #### Keyboard navigation - Present: Video controls accessible via keyboard - Improve: Add transcript navigation, keyboard shortcuts for video sections #### Contrast and readability - Present: High contrast terminal displays - Improve: Ensure sufficient contrast in any diagrams, larger text in code examples #### Text resizing and reflow - Present: Video scales appropriately - Improve: Provide resizable transcript, flexible layout for supplementary materials #### Media and captions - Present: Full transcript provided - Improve: Add closed captions with technical term clarifications, audio descriptions of visual demos #### Forms and interactions - Present: Simple link navigation - Improve: Add accessible interactive examples, keyboard-navigable configuration tools ### 🌍 DEIB Assessment #### Cultural breadth - Present: Technical content applicable globally - Improve: Add examples from different geographic regions, consider varying internet infrastructure #### Inclusive language - Present: Gender-neutral technical terminology - Improve: Define technical jargon more thoroughly, avoid Silicon Valley insider references #### Neurodiversity support - Present: Detailed step-by-step processes, clear structure - Improve: Add executive summary sections, provide multiple information processing approaches #### Representation - Present: Focus on technical merit over individual characteristics - Improve: Include diverse voices in community examples, acknowledge varying access to computational resources #### Psychological safety - Present: Encourages experimentation without judgment - Improve: Add warnings about potential costs, normalize learning through failure ### ✅ Logical Validity #### Strong reasoning - **Empirical evidence**: Demonstrates actual token consumption ($4,358 in 10 days) with specific metrics and timestamps - **Practical demonstration**: Shows working rate limit optimization and account switching in real-time - **Cost-benefit analysis**: Clear comparison between $200 subscription cost and $8,000+ inference value #### Fallacies - **Survivorship bias**: Only shows successful optimizations, doesn't discuss failed approaches or wasted tokens - **False urgency**: Creates artificial scarcity around the June 22nd deadline without considering alternative approaches - **Sunk cost reasoning**: Justifies extreme token burning because "I've already paid" rather than evaluating ongoing value #### Tone of voice - **Enthusiastic**: "This is the coolest thing ever," "Jesus Christ, we're almost at a million tokens" - **Conspiratorial**: "Take advantage of these generous subsidized subscription plans" - **Evangelical**: "Go out and build. Take advantage of these subscriptions." ### 🛡️ DARVO Analysis No DARVO patterns detected. The content focuses on technical optimization rather than defensive rhetoric. The creator openly acknowledges both benefits and drawbacks of the approaches shown. ## 🎯 Conclusion The video demonstrates sophisticated strategies for maximizing AI model access during a temporary pricing arbitrage opportunity. While the extreme token consumption shown isn't sustainable long-term, the underlying patterns of agent orchestration, workflow automation, and remote machine setup provide valuable insights for the future of AI-assisted development. Developers should experiment with these techniques while being mindful of actual value creation rather than pure token consumption. The key takeaway is approaching AI tools with excitement and creativity rather than fear, focusing on solving more problems rather than just writing more code. --- ## Full Transcript [00:00] You might have noticed I haven't been [00:01] posting as many videos the last few [00:02] days. And there's a good reason for it. [00:04] It is indeed the Fable release. It's [00:05] such a powerful model and I've been [00:06] pushing it to its absolute limits to the [00:08] point where it's affecting my sleep, my [00:09] day-to-day, and of course the content [00:11] I'm putting out and I don't want to do a [00:12] normal video where I just show some of [00:14] the cool things I've been doing with it [00:15] because this release is different. As [00:17] you all probably know by now, Fable aka [00:19] Mythos with some safeguards is available [00:21] on the pro and max tier subscriptions on [00:24] Claude Code, but only until June 22nd [00:27] because as of June 23rd, it will be [00:29] removed from those plans. The model [00:32] seems to be just too expensive and too [00:34] compute heavy for them to run in the [00:36] subscription tier that is as subsidized [00:37] as it is. And those subsidies are crazy. [00:40] We're talking like $8,000 a month. These [00:43] plans can get you up to $8,000 a month [00:45] of inference for just $200. And I'm not [00:48] talking out of my ass here. I'm speaking [00:49] from experience. I've spent the last few [00:51] days going hard on my subscriptions for [00:53] Cloud Code as well as Codex to an [00:55] extent. And I've managed to do $4,358 [00:58] of inference in the last 10 days alone. [01:01] There's a catch though. That's just on [01:02] this laptop. I also have a Mac Mini I've [01:05] been doing a lot of work on, too. And [01:07] this one's up to $1,12 [01:09] of additional inference on top of all of [01:11] this. That's a shitload of tokens. And I [01:13] want to show how I've been using them to [01:15] get real work done. And I want to be [01:17] clear, not all of this is super useful. [01:19] This is not a video about maximizing how [01:21] much you get per token or per dollar. [01:23] It's kind of the opposite. It's how you [01:24] can take advantage of these really [01:26] generous subsidized subscription plans [01:28] in the 10 days we have left with Babel [01:30] on them, which is why I'm rushing this [01:31] video out, of course, so that you can [01:33] get the most possible value and see and [01:35] taste what the future can kind of look [01:37] like with models this powerful, assuming [01:39] cost is no issue. Obviously, this is not [01:42] realistic long term. Nobody should be [01:44] spending 10 grand a month personally on [01:46] inference for day-to-day work. But [01:48] spending $200 to $400 and getting that [01:49] much inference, that is kind of [01:51] compelling. I've never burned so many [01:53] tokens in my life. In fact, I'm pretty [01:54] sure I did more in the last 10 days than [01:56] I did in the rest of my life prior to [01:58] that. And I've learned a ton since. The [02:00] good, the bad, the ugly, and more. From [02:02] loops to OpenClaw and Hermes agent to [02:04] crazy workflows that will automate value [02:07] being poured out of your code bases and [02:09] more. There's a lot to dig into here and [02:11] a lot of cool opportunities to maximize [02:13] your usage of these tools and models. [02:15] But before I can dive into all that, I [02:17] highly recommend maximizing on today's [02:19] sponsor. AI's gotten great at coding, [02:20] but deploying not so much. As great as [02:23] things like AWS are, good luck trying to [02:25] configure things properly using an [02:26] agent, especially once you're trying to [02:28] do multiple things, have preview [02:29] environments spin your team up as well. [02:31] It's just not a great story. And that's [02:33] why Render is so so cool. These guys [02:36] built an enterprise ready cloud that is [02:38] also agent ready. While every other [02:39] cloud is struggling to keep up with [02:41] what's going on, they're making nice [02:42] changes like dropping the per seat [02:44] pricing. Instead, just charging you for [02:45] your infrastructure. So, it doesn't [02:47] matter how many people or agents you [02:48] have on your team. All that matters is [02:50] how much you're shipping. You can use [02:51] Render to host anything. Web services, [02:53] databases, cron jobs, workflows, static [02:55] sites, CDN content, KVS, everything is [02:58] supported by these guys. And they have [03:00] integrations for everything you'd ever [03:01] want. And as I mentioned before, these [03:03] guys are actually enterprise ready. They [03:04] have private networks built in, so your [03:06] systems don't have to communicate over [03:07] the internet. This is huge for building [03:09] real systems. And you can configure this [03:10] all with blueprints. Simple YAML files [03:12] that described your entire [03:13] infrastructure. Imagine if Terraform [03:15] went way deeper and further. That's what [03:17] you get with their blueprints. That [03:19] said, if you want Terraform, they [03:20] support that, too. Don't worry. And if [03:21] you're building heavily distributed work [03:23] with lots of pieces with high failure [03:25] rates, render workflows are going to [03:26] save your butt. It's a simple system for [03:28] queuing and scheduling your work that [03:29] has a great SDK for TypeScript and for [03:31] Python, making it easy to make [03:33] resilient, durable workloads on the same [03:35] cloud that hosts your database, your [03:36] CDN, and more. If you're not convinced [03:37] yet, would $50 convince you? Because [03:39] they'll give you 50 bucks a credit if [03:40] you use code render Theo. Try it now at [03:42] zoidv.link/render. [03:44] Before we talk about burning all of your [03:45] tokens, I want to talk about getting as [03:47] many as possible, as in maximizing your [03:49] use of the limits that you have. As I [03:51] mentioned in my last video, I'm actually [03:53] dual wielding right now. I have two [03:54] Claude Code $200 plans that I rotate [03:57] between whenever I max one out. And I [03:59] found a lot of ways to gify that that [04:01] we'll talk about in a second. But first, [04:02] I want to talk a little bit about how [04:04] these rate limits work. This is my [04:06] second account which I haven't had to [04:08] use today which means that my 5hour [04:10] session hasn't started. This is the [04:12] first important thing to start hacking [04:14] around. You want this timer resetting as [04:16] often as possible. Realistically [04:18] speaking, I can hit the 5hour limits now [04:21] in about an hour. So if I don't have [04:23] this timer ticking and I start this work [04:25] session that will kill my limit in an [04:27] hour, then I then have to wait 4 hours [04:29] for it to reset because the timer [04:31] doesn't start until you send a message. [04:33] Thankfully, this is easy to work around. [04:35] This is all it takes. I am just in the [04:38] cloud.ai site. I'm going to say hi. I'm [04:41] going to send. I'm going to stop before [04:43] it does any real work. And now the timer [04:45] is ticking. It will now reset in 4 hours [04:47] and 50 minutes even though I haven't [04:49] dinged any of my usage. Okay, there's a [04:51] tiny tiny bit was probably used, but [04:52] basically nothing. Doing this gets you [04:55] reset as soon as possible. So now if in [04:58] 4 hours, let's say 4 and a half hours, I [05:00] start working, I don't have to worry [05:01] about hitting my limit because it's [05:02] going to get reset in 30 minutes. That's [05:04] only one of the limits you have to worry [05:05] about though because there are two. You [05:07] have the current session limit and [05:08] separately you have the all models [05:10] weekly limit which resets once a week. [05:13] Similar to the 5-hour window, this one [05:14] starts counting down after you send your [05:16] first message. Again, incentivizing you [05:18] to make sure you go hop somewhere and [05:20] trigger something that gets these limits [05:22] burning. I want to automate this though, [05:24] which you'll notice is a theme for a lot [05:25] of the things I want to show off here. I [05:27] could keep going to the website. I could [05:29] even have a browser use agent go there [05:31] for me. But at this point in time, [05:33] claude-p still counts towards your rate [05:36] limits. That will change in the near [05:37] future, which means it probably won't [05:38] work for that long, but at the very [05:40] least, the strategy will work now. So, [05:42] I'm going to set something up. This is [05:44] not meant to be like, oh, look at how [05:45] smart I am. It's meant to try and get [05:46] you into the mindset of taking advantage [05:48] of the limits that we have today. I'm [05:50] currently using Hermes agent inside of [05:52] Discord as my main like default AI agent [05:55] that does random work on my Mac Mini and [05:57] I'm going to take advantage of that [05:58] here. It already has Cloud Code set up. [06:00] So, watch and learn. I want to make sure [06:02] my Claude Code account has a recent [06:03] message at all times. Set up a cron that [06:06] triggers every 5 hours. It should run in [06:07] an empty directory. The command should [06:09] be claude-p. [06:11] And now my Hermes agent is going to go [06:13] set that up. So going forward, I will [06:15] always have something triggered on that [06:17] account to keep my rate limits moving at [06:20] all times. Also, just a pro tip I got [06:22] from Ben, I much prefer using OpenClaw [06:24] and Hermes in something like Discord [06:26] because every message gets its own [06:27] thread that I can keep following up in [06:29] to maintain context. It's really, really [06:31] nice. And I'm using this for a ton of [06:33] different stuff. Obviously, you can't [06:35] use your cloud code off directly inside [06:36] of a tool like this because Anthropic [06:38] really wants you to use claude code with [06:40] your sub, not tools like this. But you [06:42] can use your codec sub with this if you [06:43] want to minmax that too, which is what I [06:45] do. It works really good. Fantastic. Now [06:47] that that's done, I'll never again have [06:48] to worry about my limits not counting [06:50] down. Again, this isn't trying to get [06:52] more out of your plan. Like, I'm not [06:54] trying to push past my weekly limit or [06:56] anything here. I'm just trying to make [06:57] sure that you can get to 100% and get [06:59] your reset as quickly as possible. But [07:01] as I mentioned before, I'm juggling two [07:03] accounts. This might seem like really, [07:05] really annoying to do, especially if [07:07] you're using the Claude desktop app [07:08] because when you sign out and into a [07:10] different account, you lose all of your [07:11] history, all of your sessions, all of [07:13] your everything. I put out a warning [07:15] about this in the Cloud Code desktop app [07:16] for us account switching folks. And [07:19] Anthony from Anthropic actually [07:20] confirmed that they're going to fix this [07:22] behavior soon. I don't think that will [07:24] happen within the range we have Fable [07:26] for in the subscriptions. So, I wouldn't [07:28] count on this. And honestly, the CLI [07:30] still is a much better tool than the [07:32] desktop app at this point in time. [07:33] Desktop's made meaningful improvements. [07:35] I did my video roasting it, but I I [07:37] tried it a bunch the last few days. I [07:38] wouldn't recommend it. So, what does [07:40] account switching look like when you're [07:42] actually doing real work with these [07:44] tools? I could tell you how to take [07:46] advantage of the multiple accounts, but [07:47] I'd rather just show you. And when I do [07:48] this, I'm also going to show one of the [07:50] real use cases I've been doing with this [07:52] stuff. I've talked a bunch about my [07:53] recent project, Lakebed in various [07:54] videos. I'm trying to make a better [07:56] cloud for building apps that agents can [07:58] easily operate without needing to access [07:59] dashboards, deal with API keys, and all [08:01] those types of things. And I have a lot [08:03] of work in flight on this project right [08:05] now. In fact, some of it overlaps. PRS [08:08] 35, 37, and 39 are all attempts to add [08:11] file storage to Lakebed that have their [08:13] own benefits and negatives, [08:14] respectfully. 35 and 37 differed a lot [08:17] in their implementations. I had Codeex [08:19] and Claude Code both building their own [08:21] solutions and then comparing each others [08:23] and they couldn't have disagreed more. [08:25] But when I synthesized the best parts of [08:27] both and made a new PR 39, it came out a [08:29] lot better. I'm still not 100% confident [08:32] though and rather than doing the normal [08:33] thing, which is sitting there and [08:34] reading all of this yourself, I'm going [08:36] to be lazy and take advantage of my [08:38] limits. I have Cloud Code open in a [08:40] random work tree of this repo and I'm [08:42] going to tell it to get some work done [08:44] here. Notice that I'm using Ultra Code. [08:46] I don't recommend using ultra code and [08:48] the workflow feature in cloud code if [08:50] you're trying to conserve tokens at all. [08:53] But if you're trying to hit those [08:54] limits, you're trying to take advantage [08:55] of your remaining inference before a [08:56] reset comes up, doing things like this [08:58] is actually really useful and finds more [09:00] information than I would have expected. [09:02] I'm going to whisper flow this one cuz I [09:04] don't feel like typing and talking at [09:05] the same time. So forgive me. I [09:07] currently have three poll requests open [09:08] on this project, 35, 37, and 39. all of [09:12] which are implementing roughly the same [09:13] feature which is userfacing object [09:15] storage that developers can implement in [09:17] their lake bed apps. I want to decide [09:19] which one is the best choice to merge. [09:21] Make a workflow where you break these [09:23] PRs up and have judges review each of [09:25] them independently to conclude and [09:27] figure out which one is best and what [09:29] pieces of each we can bring into the [09:31] best solution. Audit all of them [09:33] independently and help me come to a [09:35] conclusion as to which of these PRs [09:36] should be the one that we continue [09:38] iterating on and merge eventually. I put [09:40] the word workflow in here specifically [09:42] to trigger it to start a workflow [09:45] because workflows are a great way to do [09:47] this type of giant bulk work. I've [09:49] actually found it very nice for this [09:50] type of thing. It's now creating the [09:52] workflow and orchestrating all of these [09:53] sub aents to be the judges. And I'll [09:55] show what that looks like in Cloud Code [09:56] because it is really cool. What I want [09:58] to show you guys first is the account [09:59] maxing that I've been doing. So here I [10:01] have my personal Claude Code account and [10:03] here I have my secondary one I was [10:05] mentioning before. The personal is the [10:07] one I currently have signed in. So, if I [10:08] refresh, you're going to see that usage [10:10] start to tick up pretty fast. [10:14] Not [10:15] okay, not that fast because it's still [10:17] figuring out what the workflow is. So, I [10:19] only have one thread going. And [10:21] honestly, running Mythos 24/7 or Fable [10:23] in this case, 24/7 in just one thread. [10:26] So, you only have one going at a time, [10:27] you're probably not going to hit the [10:28] limits too aggressively. But when you [10:30] start getting workflows going where you [10:32] have eight or more running at the same [10:33] time, you'll burn through those limits [10:35] fast, like easily under an hour. Okay, [10:37] it's still setting up the environment, [10:38] but we do see the percent starting to go [10:40] up. We're now at one. It'll go to two [10:42] momentarily. Well, it would, but I want [10:44] to demo the account swapping. It's a [10:46] really, really complex process. I'll [10:48] show you just how complex. You run cloud [10:51] code. [10:53] You run the /lo command. [10:56] You grab the URL that it puts out and [10:58] you put it in whatever browser profile [10:59] has the cloud code that you want. I'm [11:01] using Helium, which has browser profiles [11:03] just like Chrome does, so I can swap [11:04] between the two. The purple is for my [11:05] secondary account. I authorize here. I [11:08] go back to my terminal. And now I'm [11:10] logged in on a different account. And [11:13] ready for the crazy part. We had this [11:16] workflow going generating tokens. The [11:19] next turn it takes, as in the next time [11:21] it does a tool call or kicks off a [11:23] workflow or starts a next step of any [11:26] form, that one's going to use the new O [11:28] token that it just got and it's going to [11:30] start routing to the other account. It [11:32] doesn't care when you swap accounts mid [11:35] session. So if I have five workflows [11:38] going across different projects or [11:40] different work trees on the same project [11:42] and I want to make sure that I'm not [11:44] going to have them all stop immediately [11:46] because I hit a limit which is super [11:48] annoying by the way because it doesn't [11:49] recover these workflows well. So, if you [11:51] have a workflow with like 100 plus sub [11:53] aents, you get to the 94th and then your [11:56] limit hits, you have to often rerun the [11:59] entire workflow when you hit those [12:01] limits. So, it's worth it to keep an eye [12:04] on how close you're getting and swap [12:05] your O out before you hit those limits. [12:08] Or you can do what I do, which is burn a [12:09] shitload of money by turning on overage [12:12] credits. And once you see those getting [12:13] burned, then you switch over. But now [12:16] you see this account is no longer going [12:17] up because this is not the account that [12:18] the traffic is being routed through. [12:20] Okay, now the workflow is going and you [12:22] can see just how many tokens get burned [12:24] with these things. We have the audit [12:25] stage which will be 13 separate agents [12:28] and depending on what their results are, [12:30] they'll pass it on to the judge section [12:31] which will have even more than verify, [12:34] harvest, and synthesize. This run looks [12:36] like it might be 100 plus of these sub [12:38] agents, which is going to be crazy. [12:40] Burns absurd tokens. We're already at [12:41] 368,000 tokens because it's running [12:44] eight of them in parallel right now. So [12:45] now if we go here, you'll see these rate [12:47] limits for my second account are going [12:49] up super fast. We're already at 5% when [12:52] it took the last 20 minutes to even hit [12:54] 1% here because now we're running it 8x [12:56] harder. But now this account isn't going [12:58] up anymore cuz it's not the one being [13:00] used for this sub agent run. Super [13:02] helpful for hopping between accounts. [13:03] And the fact that you can just do / [13:05] login in any cloud code terminal on a [13:07] given machine and it updates all of the [13:09] things on that machine is wonderful. [13:11] Jesus Christ, we're almost at a million [13:12] tokens down and this has been like under [13:14] a minute. Yeah, this percentage is going [13:15] up fast. Like, this is real time. We're [13:17] already at seven. This is why you should [13:20] be careful with your workflows. But you [13:22] should also be careful with your weekly [13:24] limits. My two accounts reset on [13:26] Wednesday and Thursday, respectively, [13:28] right now. And I'm already at pretty [13:30] high percentages on both. I'm expecting [13:32] to max out my weekly on at least one of [13:33] these accounts, and I'm kind of counting [13:35] on a reset for some reason. Hopeful that [13:37] they'll have some reason to reset [13:38] things, but even if they don't, I'll [13:40] probably just grab another $200 account [13:42] and push it to its limits as well. [13:44] specifically during this short testing [13:45] window of 10 days where we can use [13:47] Fable. I will not be going anywhere near [13:49] this hard on anthropic models when Fable [13:51] is no longer in the subscription tiers. [13:53] But when it is, I'm pushing my limits. [13:55] So, you might be wondering now, how big [13:56] are those weekly limits? Well, from my [13:58] experience, just from basic testing and [14:00] this secondary account in particular, it [14:02] seems like you can get roughly 25% of [14:06] your weekly when you hit 100% of your 5 [14:08] hour. Put simply, you can max out a [14:11] 5-hour window four times in a given week [14:13] before you're out of usage. So, should [14:15] you always be trying to hit the 5 hour [14:17] limits? Maybe, especially during the [14:20] end, but you should absolutely be aiming [14:22] to hit the weekly limit to get the most [14:23] out of your usage. So, I've already [14:25] burned like $400 of inference just with [14:27] these basic tests as I've been filming. [14:29] And you might be thinking, "Wow, that's [14:31] a lot of inference for things that [14:32] haven't actually panned out much yet." [14:34] Like, you're reviewing three slot PRs [14:36] and trying to decide which one is best. [14:37] Like how valuable is that? First off, [14:40] that is actually quite valuable when [14:41] you're trying to figure out how to deal [14:42] with a giant pile of PRs. But second [14:44] off, you can modify this slightly to be [14:46] way more useful. Here's a much more [14:48] realistic example that I find is [14:51] actually super useful in my day-to-day [14:53] work. I maintain a handful of different [14:55] repos. from Lakebed, which still isn't [14:57] public yet, to T3 Chat, which also isn't [14:59] a public repo, but at the very least has [15:01] real people contributing every day, to [15:03] T3 code, which is a big open source [15:05] project that has a lot of people [15:06] throwing stuff at it. Even with our best [15:08] efforts, keeping it under 400 issues and [15:11] under 300 PRs, feels nearly impossible. [15:14] So rather than try to do that, I've been [15:16] putting more effort into highlighting [15:17] and surfacing the best work that is [15:19] worth pursuing. reading all those PRs [15:21] and coming up with all those changes is [15:23] way too much work for any reasonable [15:24] human to even consider trying which is [15:26] why I don't I let my agents do it [15:28] because they don't get bored. They can [15:30] just sit there and do the thing [15:31] indefinitely. So one of the things I [15:33] have Mythos doing for me, well Fable, [15:35] same difference is every morning going [15:37] through all of my PRs on all of my repos [15:40] and helping me surface the ones that are [15:41] the easiest to merge, the most [15:43] justifiable to get done, etc. This [15:45] example is using my Hermes agent on top [15:47] of my codec sub, but this is the one [15:49] that Mythos made and it's really good. I [15:52] ran this one on just T3 code and it [15:54] built this ranked Q section where it [15:56] goes through every PR that's currently [15:58] open, gives it a status and ranks them [16:00] based on how easy is this to just go [16:02] merge and how much of my attention does [16:04] it deserve. So number one here we have [16:06] disabling external git diffs, which is a [16:08] PR that Magnus opened to fix a bug with [16:10] the external diff viewers some people [16:11] use with git when they're using T3 code. [16:13] Simple PR fixes a real bug. People will [16:15] be very happy if we merge this. It was [16:17] surfaced by the agent, even though this [16:19] PR was originally filed over a month [16:21] ago. And as soon as I sent this link [16:23] with all of these ranked PRs to Julius, [16:25] it got merged within 5 minutes. Getting [16:27] this type of overview of all of your [16:29] work is so much better than trying to [16:30] dig through the GitHub PR tab trying to [16:33] get useful information out of it. If [16:34] you're wondering how I got my agent to [16:36] spit out these HTML plans, that's [16:38] another service that I threw together. [16:39] This one's just for me and my team, but [16:41] I recommend building something like it. [16:42] And maybe I'll even open source it in [16:44] the future if enough people want it. I [16:45] built a simple service for hosting plan [16:47] files, which are just HTML plan [16:49] descriptions on a real web service so my [16:52] agent can spit out a URL that I can [16:53] click on and then go see what it was [16:55] thinking or what it's planning. I find [16:56] HTML plans to be way more readable than [16:58] markdown. I have a whole video about [17:00] this. And I've been abusing these for [17:02] all sorts of different things, including [17:04] reviewing agent work and having one [17:05] agent review another agent's work and [17:07] then just pass the HTML over, but also [17:09] having the ability to read it in a good [17:11] format. That's what's cool about HTML is [17:13] I can look at in my browser or on my [17:14] phone in my browser and get a good idea [17:16] of what's going on and then just paste [17:18] the link to an agent and say, "Hey, go [17:20] deal with this." And it figures it out. [17:21] By the way, that agent run I started [17:23] earlier with the workflow is now at 1.6 [17:25] million tokens and 21% of this fresh [17:28] usage window used. And that is in under [17:32] 30 minutes. Kind of crazy. You can burn [17:35] through usage fast, but if you sear it [17:37] in the right way, you can get actually [17:38] useful stuff. I actually did have PR39 [17:42] get reviewed previously by Mythos and I [17:44] have its review here which is super [17:46] convenient because Codex was actually [17:48] the agent that wrote this PR. So having [17:50] Codex write it, having Mythos review it [17:52] and then I can take this URL, hand it to [17:54] Codeex to say, "Hey, go make the changes [17:56] that this suggested and give me feedback [17:58] on the ones you don't like as much, it's [18:00] such a useful way to just pass context [18:02] around in a way I can read." So yeah, [18:04] again, points in favor of HTML plans. [18:06] Highly recommend it. If you're already [18:08] pushing the limits of workflows and [18:09] reviews and you're still not getting [18:11] close to those limits, there are plenty [18:12] more things you can do, don't worry. One [18:14] that I would highly recommend is looking [18:16] into the skills others have made to [18:17] audit your projects and your code and [18:19] find ways to improve it. For example, [18:21] Shad CN's Shad CN improve skill is [18:23] really nice. Versel built this nice [18:25] plugin for adding skills. If you want to [18:27] actually add the skills, there's cool [18:28] commands like the Verscell skills [18:30] package, but you can also just go copy [18:31] the content of the skill directly [18:33] because as you hopefully know by now, [18:35] skills are for the most part just [18:37] markdown. [18:39] So here we have all the markdown for [18:41] this skill. I can go to raw. I can grab [18:44] this all. I can open up claude. I could [18:47] put it in workflow mode or ultra code [18:49] mode, but I am already burning enough [18:51] inference and I want to save some of [18:52] this for later. So, I'm just going to [18:53] paste enter and let it do its thing. [18:56] Again, very useful if you notice [18:58] yourself at the end of a weekly limit or [19:00] an hourly limit and you want to get a [19:02] little bit more inference out. Just keep [19:04] a set of these things in your head that [19:06] you would like to run and write them [19:07] down. Maybe even cue them. I've also [19:10] recently been loving Raycast's notes [19:12] feature. It's super easy to just like [19:14] write something down [19:16] and then hide and reopen it with a [19:18] hotkey. And you can easily iterate [19:20] through the notes that you have saved [19:22] here to find random prompts or things [19:24] you want to do and go grab them in the [19:26] future if you don't want to send them [19:27] just yet. Looks like it's going to kick [19:29] off a workflow for the improving [19:31] anyways, which means I'm about to burn a [19:33] lot of money. Great. Thankfully, it's [19:36] not my money. And since starting this [19:38] video recording, we've already done [19:39] another like $400 of inference. Jesus [19:41] Christ. Take advantage of these generous [19:43] limits while we have them, folks. We got [19:44] to escape that permanent underglass. One [19:46] thing I touched on earlier that I [19:48] haven't really dug into yet is the [19:49] effect this has had on my sleep. I'm [19:51] sure we've all been there. Glued to our [19:53] laptop just waiting to get the result so [19:55] you can run the next prompt and then go [19:57] to bed and then the one more prompt [19:59] effect keeps you going and not leaving [20:01] your desk again and again until [20:02] eventually you're like falling asleep at [20:04] your keyboard. The AI vampires of the [20:06] Silicon Valley is a trope that I'm [20:08] seeing more and more. Even some of the [20:09] most skilled maintainers that are [20:11] fathers of loving families are finding [20:13] themselves staying up till 4 in the [20:14] morning prompting because it's so [20:16] addicting. I don't have a solution to [20:17] that. I'm not gonna sit here and pretend [20:18] I do. What I do have is a Mac Mini on [20:21] the same network that has Codeex, Claude [20:22] Code, and Hermes Agent all set up and [20:24] ready to go on it, which has been [20:26] wonderful for being able to get [ __ ] [20:28] done from my phone and from other [20:30] computers. And most importantly, being [20:32] able to close the lid on my laptop and [20:34] not worry about the work stopping. I [20:36] have three ways I interface with that [20:37] Mac Mini. the one that I admittedly use [20:39] the most is the SSH directly into it. [20:42] When I'm on the same network, it's very [20:43] easy to SSH into. When I'm not, I rely [20:46] on Tail Scale to do it, which has been [20:47] much better than I was expecting. As an [20:49] old WireGuard fan, seeing Tail Scale in [20:51] its current state is awesome. Highly [20:52] recommend. I highly recommend using a [20:54] terminal like CMUX, which has both the [20:56] sidebar and tabs, so you can pin your [20:59] Mac Mini or whatever other computer you [21:00] have over SSH as its own section and [21:03] easily manage and navigate it. It's been [21:05] super nice. Since I'm on a Mac and the [21:06] remote computer is a Mac, I can use the [21:08] built-in screen share utility, which [21:10] honestly I didn't even know about until [21:11] recently, which makes it super easy to [21:13] access the computer, have native [21:14] hotkeys, and all the other things you [21:16] would expect when you're using this [21:18] computer remotely, which has been again [21:20] a lifesaver for doing this type of [21:22] thing. This also works over tail scale, [21:23] which let me configure some things that [21:25] were blocking when I was away from my [21:27] computer. But none of those are my [21:29] favorite way to access that remote [21:30] machine. And this is where I will admit [21:32] we're about to get into a bit of a [21:33] selfplug, but I hope you guys appreciate [21:35] it because I think this is the coolest [21:36] thing ever. Julius has put a ton of work [21:39] into the T3 Code remote system where you [21:41] can control a T3 Code instance from [21:44] another computer. I usually do it by [21:46] just connecting directly to the machine [21:48] that I have T3 Code set up on, but [21:50] there's lots of other options, too. If [21:52] you have T3 code installed on one [21:53] machine, it's really easy to give access [21:56] to other ones. [21:58] You can create a link or expose it over [22:00] something like tailscale and then access [22:01] it through that link or by going to the [22:04] app.t3.code site and creating the [22:07] connection there. Julius is also deep in [22:08] the building of T3 connect which is a [22:10] new method to connect to a remote T3 [22:13] code instance that you have on one of [22:14] your machines from another machine or [22:16] even your phone with the upcoming T3 [22:17] code app. Very excited about everything [22:19] we're cooking there. But I'll be real, [22:21] you can build all this yourself. Codex [22:23] has some of it built in already. Cloud [22:24] code pretends to, but it never works [22:26] when I try. There's lots of ways to [22:28] control your agents remotely. And I [22:30] highly recommend finding methods like [22:31] this because you'll start pushing the [22:33] length of your jobs a ton when it's no [22:36] longer locking you to your laptop. [22:37] Before I had this remote Mac Mini set [22:39] up, I found myself running shorter jobs [22:41] so I wouldn't have to worry about [22:42] closing my laptop cuz I didn't want to [22:44] be one of those people walking around [22:45] with a halfopen laptop. I really didn't. [22:47] Once I had the Mac Mini set up and I [22:49] could run things from there, I found [22:50] myself using agents entirely [22:52] differently, letting them go off on long [22:54] exploratory journeys, not expecting to [22:56] be able to use the results, but [22:58] expecting to have interesting enough [22:59] findings to talk about to my agents or [23:02] even to my other co-workers when we [23:04] build these types of things. I don't [23:05] care what method you end up using to [23:06] control these remote machines. You can [23:08] even build your own, which is a fun way [23:09] to both burn tokens and really refine [23:12] your workflow to fit your specific needs [23:14] and expectations. If you take anything [23:16] from this video, I really want you to [23:18] take home the creativity you can apply [23:20] to how you burn your tokens. Now, you [23:22] can have agents reviewing each other's [23:24] work, and you can even automate it. [23:26] Something like Claude Code is capable of [23:28] doing babysitting with built-in loops. [23:30] So, you can tell one agent to make a PR [23:32] and watch whenever a new comment comes [23:34] in, and then tell another agent to watch [23:36] a PR and whenever a new push happens, [23:38] give it a bunch of feedback and leave a [23:40] review. Now you have two agents that [23:42] aren't even aware of each other giving [23:44] each other what they need to keep [23:45] pushing work forward before a human's [23:47] even involved. You can give agents stuff [23:49] like browser use. This is much stronger [23:51] on the codec side, but Cloud Code is [23:52] starting to get there as well. This [23:54] allows an agent to spin up the changes [23:55] it made, record the screen showing that [23:58] the changes work, and then send you that [23:59] or post it in the PR when they're done, [24:01] not bothering you until they have video [24:03] evidence that the work actually worked. [24:06] As always, Pete is far ahead of the game [24:08] and he has awesome examples. Codeex has [24:10] the ability to spin up threads by [24:12] itself. So it's not just sub ages. It [24:14] can trigger new threads and make new [24:17] work trees and make real changes through [24:19] that. So you set up a really simple loop [24:20] here. Tell Codeex to maintain your [24:22] repos. Wake up every 5 minutes and [24:23] direct work to threads. Makes it easy to [24:26] parallelize and steer work as needed. [24:28] Since each thing that it thinks should [24:30] be going on gets its own thread, it's [24:32] very easy to hop into a given thread and [24:34] tell it, "Hey, that's wrong. Go do it [24:36] the other way or just stop it if it [24:38] doesn't make sense." If you combine this [24:39] with something like your sentry bug [24:40] tracking or your data dog analytics, you [24:42] can combine it and give all of the [24:44] resources and data available to you as [24:46] the engineer to your agents, which let [24:48] them find information that's useful, [24:50] self-improve, and just push things [24:52] forward in a way that's really powerful [24:53] and surprisingly valuable. Jesus Christ, [24:56] that workflow is now at 1.8 million [24:58] tokens. We haven't even started the [24:59] judges yet cuz we got one more of these [25:01] audit agents going. it. I know it's like [25:04] I'm lighting money on fire here, but [25:05] I've already paid the money and it's [25:07] actually really fun to watch. I didn't [25:09] think I'd ever be this type. I've always [25:10] been the like singlethreaded just get [25:12] the thing done guy. I'm having a lot of [25:14] fun right now. I've been going a little [25:16] further than I normally do, and I really [25:18] like how Sawyer framed this. You need to [25:20] be more ambitious than you have been [25:21] before. Ask the model to rewrite your [25:23] entire production app from scratch. Ask [25:25] it to make an entire product or internal [25:27] tool for you, but don't stop there. Ask [25:29] it to deploy it and add accounts, [25:31] multiplayer, etc. All the things you [25:33] wouldn't normally have in a throwaway [25:34] personal app. Raise your bar. Only by [25:37] pushing more do you learn. Orchestrate. [25:39] When coding and having an agent write a [25:41] plan, have another agent or multiple. [25:43] Validate each claim in that plan. Quad [25:45] workflows are [ __ ] amazing for this. [25:46] Sub agents and codecs are getting there [25:48] now, too. They're not quite as good, but [25:49] they're better than I thought they would [25:50] be. Fable's really good at [25:52] orchestration. One tip with that though, [25:54] if you don't want to burn too much, [25:56] explicitly tell Fable during [25:57] orchestration to use Opus or Sonnet for [26:00] its sub agents because for whatever [26:02] reason, Fable more than any other model [26:04] really likes using Fable for its sub [26:06] agents. It seems to not trust other [26:08] models. So, be explicit. Tell it to do [26:10] other things. Also, see there, it just [26:12] finished the audit stage and then it [26:14] reviewed what it got back and decided it [26:15] needed seven judge agents to do more [26:18] work there, as well as these harvest [26:20] agents that spun up in the background as [26:22] well. But again, Fable's incredible at [26:24] orchestration. Fable is not needed for [26:26] the individual sub agents. If you're not [26:28] trying to actually burn as much money as [26:30] possible, maybe tell Fable to use [26:32] something else. Back to what Sawyer said [26:34] here, have it use Codex-P to delegate [26:36] parts of it to codeex. Have it use Opus [26:38] sub aents. Have it use workflows. Don't [26:40] be afraid to give those models big [26:42] chunks of work. Give your agent a [26:44] browser and computer use. Let them click [26:46] around to take screenshots. Use the JS [26:47] debugger and take performance profiles. [26:50] LMS aren't just coding machines. They [26:51] can do the debugging and reproducing [26:53] work as well. For token maxing, get some [26:56] way to kick off coding tasks from your [26:57] phone. Could be the Codex app, ChatgBT, [26:59] Claw, etc. As I mentioned before, I [27:01] threw a bunch of this in my Discord with [27:03] my Hermes agent. Whatever you choose to [27:04] do, doesn't matter. Make it easy to get [27:07] the idea out of your head into an agent [27:09] running. I don't care how you set this [27:11] up. There's lots of good options, but [27:12] your goal should be to make it as easy [27:14] as possible to go from, "Oh, that's a [27:16] nice idea," to having something running, [27:18] testing out that theory for you as [27:20] quickly and smoothly as possible. The [27:22] more of these opportunities you take, [27:24] the more of them you'll find, and the [27:25] more you find, the more you can do, and [27:27] the more you'll find the limits of the [27:28] tools that we're relying on now every [27:30] day. My last thoughts about where this [27:32] is all going, and more importantly, how [27:34] we should feel about it. The amount of [27:36] work somebody can do as a developer has [27:38] gone up exponentially as a result of all [27:40] of this. Any one dev can do 10 times to [27:43] 100 times more work than they could [27:45] before. They can't necessarily validate [27:47] that that work is good or worth [27:48] shipping, but they can absolutely get [27:50] the code out more than they ever could. [27:52] Your focus should not just be writing [27:53] more code. Your focus should be solving [27:56] more problems. And you shouldn't be [27:58] doing this out of fear either. If you [28:00] come into this scared that you're going [28:01] to lose your job if you don't token max, [28:03] you're going to burn a bunch of tokens [28:05] and a bunch of your own sanity and come [28:07] out way more stressed and not happy and [28:09] probably still unemployed. If you go [28:11] into this with excitement, the [28:13] excitement that we all went into [28:14] software with, with the ability to [28:16] customize our computers and the things [28:18] on them to do what we want and need [28:19] specifically for ourselves. If you go [28:21] into your agents with that mindset, if [28:23] you go into the token maxing with that [28:26] goal, trying to get more out of the [28:28] things that you do and build every day, [28:30] if you approach it with excitement, what [28:32] you'll get back is incredible. When I [28:34] went into this stuff skeptical, what I [28:36] got out was some okay code and a bunch [28:39] of bugs. When I went into these things [28:40] excited to push the limits of what I [28:43] could do and build, I came out with my [28:45] own custom cloud. I came out with a pile [28:46] of npm packages I use every single day [28:49] for a ton of different stuff. I started [28:51] forking the software I rely on to tailor [28:52] it to my specific needs. I started [28:54] building my own control planes for [28:56] managing all of this. I started having [28:58] way more fun and building way more, too. [29:00] So, don't approach this with the [29:02] attitude of keeping your job and making [29:04] sure you'll still be employed in a few [29:05] years. Approach this with the excitement [29:07] that we can build all the things we ever [29:09] imagined. Lower your bar for what's [29:11] worth building and raise your bar for [29:13] how far you bother going. And what [29:15] you'll find is a lot of fun, but also a [29:18] lot of rate limits being hit. We're at [29:19] 42% now and we still have four hours [29:22] left. God. [29:24] Yeah, I'm voted by two subs and I think [29:26] you should too. Mythos is an incredible [29:28] model capable of things I never would [29:30] have expected even a year ago for agents [29:32] to be able to do. It builds awesome [29:34] things. It writes awesome plans. It [29:36] describes the stuff it's doing in ways [29:37] that are actually digestible. And you [29:39] can use this to push the limits of what [29:40] your code is doing to get better [29:42] information and spend less of your time [29:45] doing the things you don't like and more [29:46] of your time shipping things you do. If [29:48] you find yourself stopping to ask, can [29:50] an agent do this? Reset your mind. Just [29:53] ask the agent to do it and see if it [29:55] fails. And if it fails, try to figure [29:57] out why. Ask it what it struggled with. [29:59] Look through the history and the bad [30:00] tool calls it made and try to steer it [30:02] in the right direction. Maybe make a [30:04] skill that points it in the right way. [30:06] Maybe build your own verification in [30:07] your codebase to push it in the right [30:09] direction. If you're not already writing [30:10] your own custom lint rules for weird [30:12] [ __ ] that your agents are doing, you're [30:14] not being creative enough yet. And [30:15] that's fine. This is a different way of [30:17] thinking and it's a really exciting one. [30:19] So go out and build. Take advantage of [30:20] these subscriptions. Push the limits of [30:22] the best models and tools available [30:24] today and don't get too locked into them [30:25] because new things will be available [30:27] tomorrow. Just try to be excited when [30:29] you go in and you'll be amazed what you [30:31] can pull out. Now that I got these [30:32] agents cued, I'm going to go catch up on [30:33] some sleep. I recommend you do the same, [30:35] but come back fresh and build some cool [30:36] stuff. Let me know what y'all are [30:38] building.  📄 **Full text:** ![[Full Document Contents/Articles/Mythos Is Here, It’s Time to Start Tokenmaxxing.md|Mythos Is Here, It’s Time to Start Tokenmaxxing by YouTube]] ## Highlights - you all probably know by now, Fable aka Mythos with some safeguards is available on the pro and max tier subscriptions on Claude Code, but only until June 22nd ([View Highlight](https://read.readwise.io/read/01ktxwzgmtkm6d4x6yydw2sd67)) - Note: Fable (= Mythos + safeguards) on Pro/Max Claude Code subscriptions, but only through June 22 - Tags: atom/literature/h1, tokenomics - plans can get you up to $8,000 a month of inference for just $200. And I'm not ([View Highlight](https://read.readwise.io/read/01ktxwzj9et0nx8yg8zd6971at)) - Note: The "tokenmaxxing" arbitrage: subscription plans subsidize up to ~$8,000/mo of inference for $200 - Tags: tokenomics, atom/literature/h2