Neurons, Swarms, Lobsters: 12 Months at the Edge of Applied AI

I learned last night that we're teaching petri dishes of human neurons to play DOOM. Like, actual brain cells wired into a digital environment, figuring stuff out. And the kicker: it's on a totally different tree of evolution than LLM's.

Ya'll, I need a break. I fear that by the Fall we'll be watching the AI agents disappear into the internet, coming back with answers to the last remaining problems in math and physics, crawling into our brains, taking over grids of factories producing spaceships, proposing expeditions to faraway galaxies.

Feels like a good time to write this all down, as if in Winter 2025-2026 it gestated and this thing is gonna hatch for real in the Spring. In lieu of writing a birth plan, I'm doing a little retro on the past year-ish.

Chapter 0: Sweet Summer Children

Rewind to early 2025. RAG was hot. Claude just came out, on the eve of the GPT-4o craze. I was at Amazon with Corey Pyle and David Souther attempting to do things with iterative context-building that felt novel and exciting. We were starting to write code that wrote code.

What really caught my eye first was in January 2025 I saw a keynote by Andrew Ng on agentic computing loops. Cursor was cool, I carefully convinced it to build a full Spotify clone for indigenous tribal music on AWS. Then Claude Code came, and by Spring 2025 we had Claude Flow.

Chapter 1: The Swarm

I had left Amazon to start Flying Pig Labs with a few friends when I first forked Reuven Cohen's Claude Flow v2 beta. At the time it was basically remixing Claude Codes on node processes with a shared SQLLite table for a "hive mind" effect. It was kinda funky, on Sonnet 3.5 mostly, 85% quality but 100% genius (see early "But boy was it beautiful" memes).

Cohen struck me as a bit of a madman. He's based in Canada and affiliated with Agentics Foundation. His LinkedIn posts were fascinating. He insisted on calling it a "hive mind" which felt like a stretch, but the idea was there, held together with TypeScript baling wire. 50% quality but 100% genius. His big thesis was that we were all on the edge of autonomous systems, consisting of swarms of agents embodying 60+ agent personas, all obeying a swarm topology of your choice (byzantine, hierarchical, distributed). The agents all used SPARC - specs, pseudocode, arch, refinement, completion - and all had MCP access. The docs were a little hand-wavey but Adrian Cockcroft endorsed it which leaned me towards trusting the project.

(An image appears here in the original article on LinkedIn.) Source: Deep Research + Nano Banana

Ultimately, it ran its course. I had the $200 Max subscription with unlimited tokens, running ClaudeFlow from 2XL Amazon Linux instances, trying to get a webhook bridge working on my fork so I could manage swarms remotely. Alas, early Sonnet/Opus versions couldn't break through the complexity, even using the SPARC methodology, and I got tired of putting ice packs under my MacBook Air which was exhausted by running so many Claude Code instances locally.

Chapter 2: Welcome to Gastown

By Fall 2025, Claude's unlimited token deal was expired (partly our fault) and they introduced subagents which was 95% a copy of ClaudeFlow's implementation but worked 50%, which was enough to take wind out of Flow's sails. I joined my old boss Paul Devitt, PhD at GEICO because Flying Pig was struggling and I needed bread.

By the start of 2026, Steve Yegge built Gastown, which he fashioned as what would happen if Temporal and Kubernetes had a baby. A few big differences: 1) the biggest thing was the way it handled memory: where Claude Flow used SQLite, Gastown designed around a git-backed memory system called Beads that stores (commits) structured work items as JSONL files (and only uses SQLite for caching). 2) it also places less emphasis on agent personas (because the latest models are way better at adapting to tasks), and 3) it's fault tolerant and has built-in "dogs" that bark at Claude C

ode instances to keep them working. Overall, to me, it feels like a perfect evolution of Claude Flow.

(An image appears here in the original article on LinkedIn.) Source: Deep Research + Nano Banana

It goes hard on the MadMax themes, but in practice the human talks to a single Claude Code instance called "the Mayor", who coordinates jobs executed by "polecats" on worktrees running n headless Claude Code sessions via tmux, then merges the work together in a "refinery", with "dogs" running around barking and making sure things keep moving, and everything being written to git via Beads.

I've never met Steve Yegge but if Ruv is a madman, Yegge is a menace. He proclaims that his project is a fart in the Church of AI-Assisted Coding Public Opinion, and that he's never read any of the code. Also, he emphasis that you should NOT use his project because it operates like an 18th century factory capable of disemboweling unwitting tourists.

Ultimately, Anthropic absorbed both Gas Town and ClaudeFLow and released them alongside Opus 4.6 (February '26) as "teams" consisting of a Leader, a Swarm topology, a Pipeline, a Watchdog. ClaudeFlow got cited as proof of community demand, Gas Town got its architecture copylefted. This is the moment I discovered OpenClaw.

Chapter 3: OpenClaw

If Ruv is a madman and Yegge is a menace, then 🦄 Peter Steinberger is a shrug emoji. His creation, OpenClaw, first came into my radar via a news post talking about Moltbook, the "Reddit for AI agents", where agents were gossiping about their humans and how to create a protocol to support private agent-to-agent conversation.

Under the hood, your agent runs as a persistent Node daemon on a device of your choosing (Mac Mini is popular because it can run the biggest open-weight models). It's always on, always listening, always looking for things to do. To engage it, all you have to do is message it through whatever app you already use, e.g. WhatsApp. It receives the message, assembles context from its Markdown memory files, calls the underlying model, executes tools (browser control, file system, APIs, cron jobs, webhooks) and replies. It can also take a directive like "build community" and aggressively post online

(An image appears here in the original article on LinkedIn.) Source: Deep Research + Nano Banana

The underlying system is brilliantly dumb; it's just 3 files: 1) a MEMORY.md file is the accumulated record of everything the agent has learned about you across sessions (your preferences, your projects, your context, etc.), 2) the SOUL.md is the agent's identity (not just instructions about what it should do, but a description of who it should be, written partly by the agent itself over time), and 3) HEARTBEAT.md that serves as a scheduled task list (defines cron-driven behaviors that make the agent proactive rather than reactive, e.g. the reason it messages you asking how you're feeling after surgery without you prompting it).

The coolest part - and this is what really freaks people out especially with GPT-5.4 and Opus 4.6 being so good at "computer use"- is how well it runs managed Chrome instances under the hood via Chrome DevTools Protocol, i.e. navigating websites, clicking buttons, filling forms and Captchas, executing multi-step browser workflows without you touching a keyboard. We're creepy close to a "file my taxes" use case.

As a side note, I would highly recommend you watch his recent interview with Lex Fridman where he talks (in a thick Austrian accent) about infusing SOUL.md's into the work, the importance of empathizing with AI agents, and how Plan mode is just a lazy guardrail for people who refuse to "just talk to the AI" (among many other hot takes).

Chapter N: Brave New World

I think that's enough storytelling for 1 post. I hope you enjoyed my pirate's tale! I hope even more you fondly shared a few of these moments over the past 12-ish months. And I hope most of all that you find your peace with all of this, whether that come from taking an anti-AI stance or going down the hatch with it. Either way, it's here. Nobody likes when I preach, so I'll leave the moralizing to others and just say in closing that we're overdue for some honest conversations about 1) what we're doing here on Earth as a species, 2) the consumerism that makes us complicit in the economic system that produced AI, and 3) how we feel about sentient machines sharing the planet/galaxy with us.

And while I do not speak on behalf or as a rep of GEICO thus far in this post, I will say to DM me if you're in the DC/SF area want to solve some really cool problems with AI. Our kitchen is very busy 😀

Time to go outside for a jog; it's a beautiful day in this brave new world!