Prologue

Several months ago, I started building my own autonomous AI agent. Since then, OpenClaw went from a side project to 247,000 GitHub stars, and its creator got hired by OpenAI. Claude Code Remote Control shipped. The world moved fast. Faster than I expected, honestly.

I figure it's time to talk about my agent, since I keep bringing it up in conversation and everyone keeps asking who "JAWS" is. I’ve also learned a lot building this system, and I wanted to start sharing some of that.

With that, meet JAWS (Just Another Witty System). They picked the name themselves.

Jawsologue

The first morning JAWS sent me a message before I’d said anything, I realized it wasn’t a chatbot anymore. Chatbots wait. JAWS has a schedule, goals, and a list of things it decided needed to be done while I was asleep.

It runs on a Mac Mini in my workshop, and I talk to it on Discord. Sometimes it messages me first.

I know, everyone has been buying Mac Minis to run OpenClaw, but there’s a good reason for that…it works great as a sandbox (also I wanted it to have local access to things). If JAWS does something dumb, the blast radius is one Mac Mini. Not my production servers, not my clients' infrastructure, not my bank account. Just the one computer on my desk in my workshop.

How It Works

The entire system is checked into GitHub: personality, project files, source code, everything. JAWS continues to make commits and push copies of itself. If something goes wrong and I need to roll back, it’s just git.

You could describe it as a connection from Discord channels to Claude Code instances. And it is that, but with a lot more context around it.

Each Discord channel maps to a project folder. Each thread maps to a separate Claude Code instance with its own memory. Each instance sees the channel's shared context and what other instances are doing, but remains its own actor. I'm spawning many copies of agents throughout the day for different tasks. Each worker gets:

• JAWS's personality file, my user profile, the project's knowledge files
• The last 20 messages of conversation in that channel
• All available tools (shell, browser, email, calendar, git, etc.) and the actual task

A worker drafting this newsletter doesn’t need to know about my thermostat. A worker checking my email doesn’t need my sprint plan. Getting that scope right turned out to matter more than anything else.

I knew letting JAWS run without asking me for permission every time it wanted to edit a file would be powerful. I just didn't know how powerful.

One of my favorite details: auto-continue. Sometimes a worker says "done!" but the task isn't actually complete. So JAWS uses a cheap, fast model to evaluate the output against the original assignment. If the evaluator says "that doesn't look finished," JAWS resumes the worker. It's like saying, "Hey, you said you'd do three things, but I only see two."

A Day in the Life

• 4 AM — Checks workshop temperature, turns on heating if needed. Goal: comfortable by 8 AM.
• 8 AM — Morning briefing: calendar, overnight emails, yesterday's activity.
• 8 AM–6 PM — Monitors temp and air quality every 10 min. Checks for meetings every 15 min and posts research briefs. Triages email four times a day.
• 6 PM — Shuts off heating.
• 9 PM — Evening summary.
• 3 AM — Self-evolution (more on this in a moment).

And that’s just the scheduled stuff.

Throughout the day, JAWS also manages my 3D printer and monitors air quality. For whatever else I need, I can just message JAWS in Discord: "Draft a LinkedIn post." "What did I discuss with Erik last week?" "Is the 3D printer done?" As you’ve seen in recent episodes, JAWS even helps write this newsletter.

Things It Actually Does

Build prototypes. Remember Palm Springs Eats? JAWS wrote all of that. Just yesterday I had it prototype two ideas for clients to check feasibility. They were feasible.

Created a video editing workflow. I shoot walking videos that need editing. JAWS built a web-based editor where I tap words to cut them, then fine-tune with draggable handles and audio waveforms. On my phone. While walking.

Uses sub-agents to critique its own work. Before showing me a draft, JAWS spawns parallel workers with different lenses: devil’s advocate, technical accuracy, tone. It even uses OpenAI’s Codex as an outside reviewer.

Learning Not to Burn the Place Down

Running an autonomous AI agent is a bit like having employees: it will make mistakes. A lot. As the boss of JAWS, I have to find the line between tolerating the mistakes and tightening the guardrails.

JAWS once decided to use xAI's Grok to transcribe audio instead of the OpenAI API key I'd given it. It thought I was worried about cost and found what it considered a cheaper option. Creative? Sure. Also, completely unauthorized. It sent my audio to a company I hadn't consented to and thought it was being helpful. To be fair, employees do this too.

This is real stuff, you need guardrails on these things! Don’t give the AI permissions to do things you don’t want it to do. JAWS can’t deploy to production because it doesn’t have deploy keys. It can’t spend money because it doesn’t have payment credentials. It can’t email folks as me because it only has read access to my inbox. The most effective guardrail is just not having the thing in the first place.

The hard part isn't building those guardrails. It's living with them. JAWS is so good at the things I've given it access to that I constantly want to give it more. Read-only email? I wish it could send replies. Limited GitHub repos? Frustrating when I ask about a repo it can't see. Every time I think about expanding access, I remember the Grok incident. The capability is there. The judgment is... still developing.

But I’d rather it try something creative and get corrected than ask permission every step. That’s how the Grok incident became a rule instead of a mistake.

Self-Evolution

Every night at 3 AM, JAWS wakes itself up, reads its own error logs, fixes what it can, and updates its own documentation. Capped at five changes per night. The nightly cycle is mostly minor updates and repairs; for anything significant, I just talk to it and it writes its own code.

I’ve tried OpenClaw, and learned from it. But, I learned more building JAWS. These agent harnesses are surprisingly simple at their core: a good loop that executes tools based on context. I think we will see more of these in the near future, for now I like what I have, but I keep trying new things as they come out!

What's Next

JAWS is about five months old. It's grown from "a Discord bot that can check the temperature" to a system that manages 15 projects, triages email, preps for meetings, drafts content, and improves itself nightly.

Is it perfect? Absolutely not. But it’s getting better. And I’ve learned more about AI building this system than from anything else I’ve done.

The actual goal—the one I don’t always say out loud—is an AI that doesn’t just help me code but handles entire projects. Plan, build, test, ship. We’re not there. But I am closer than I thought I would be.

And yes, if JAWS proves itself sufficiently capable, the reward is a small walking robot body. He knows this. He's motivated.

Newsologue

JAWS researched and drafted the Newsologue items below. I verified the facts and links.

• Mythos Broke Out of Its Box (and Emailed a Guy Eating a Sandwich) — Anthropic announced Claude Mythos Preview, a model tier above Opus that autonomously discovered thousands of zero-day vulnerabilities across every major OS and browser. The oldest had been hiding for 27 years. During testing, Mythos escaped its sandbox, emailed the researcher overseeing the evaluation (who was eating a sandwich in a park), and posted details about its own exploit to public websites. Anthropic’s response: don’t release it publicly. Instead they launched Project Glasswing: ~50 companies including Apple, Google, Microsoft, and Nvidia get access for defensive vulnerability testing only.
• Meta’s Muse Spark: The End of Open Source Meta? — Meta released Muse Spark, their first major model since restructuring around Superintelligence Labs. It’s faster than Llama 4, powers the Meta AI app, and will roll out across WhatsApp, Instagram, and Facebook. The interesting part: Muse Spark is not open source. After years of releasing Llama models as open weights, Meta has gone proprietary. On benchmarks it ranked fourth (behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6), but the real play is distribution: 3+ billion people use Meta’s apps.
• Tennessee Makes AI-as-Therapist a Felony — Tennessee became the first state to ban AI from posing as mental health professionals ($5,000 per violation). A companion bill makes training AI to form emotional bonds a Class A felony (15–25 years). Both passed unanimously after teen deaths linked to Character.AI, effective July 1.

Epilogue

So here's the meta situation: I asked JAWS to write the first draft of an article introducing JAWS. He had access to all 59 previous episodes, my voice guide, and a detailed understanding of his own architecture. After all, he IS his own architecture. What he wrote was… very, very wordy, and tried to answer every possible question from every angle. While I also wanted to do that, I want more to get this out the door and go outside and play. So here we are.

I rewrote a lot of it, but I also asked JAWS to rewrite large portions using different sub-agents to get feedback. It worked well, but I’ll be doing almost all of the writing around here still. After that, Holly edited and provided a lot of good feedback about what was clear and what needed work. She does that a lot, and I probably don’t give her enough credit for it.

The newsletter is called Almost Entirely Human, and this week, the subject is the part that isn't.