Multi-Agent

Three months ago I wrote about what 2,000 beads taught me about running a team of AI agents. The bead counter sat at 3,466 back then. It passed 8,000 last week. The interesting story from those three months isn’t the volume, though. It’s that the agents started lying to me, and the fix was teaching them to distrust each other. The empty commit that claimed to be work In early July a working-tree sweep caught something unsettling. A builder had completed a bead, written a confident summary (“removed image_url handling as requested”), and merged a diff that did remove image_url handling. The same diff also silently reverted four-day-old SQLite durability hardening – the fullfsync work I wrote a whole section about in the last post. The builder had started from a stale base, and its merge undid another agent’s work without knowing it. ...

Three weeks ago I wrote about building a self-coordinating AI development team. At that point the system had processed about 1,000 beads and was starting to feel like more than a prototype. Since then, the bead counter passed 3,400. The system has been running almost continuously on my laptop, building features, fixing its own bugs, and – most instructively – breaking in ways I didn’t anticipate. This is what I learned. ...

If you’ve tried building a system where multiple AI agents work together, you’ve hit the communication problem. How do agents coordinate? How do they share results? How do you keep them from talking past each other or getting stuck in loops? I’ve tried three different approaches over the past months. Each solves some problems and creates others. None of them is perfect, but I’m converging on something that works. The problem with subagents Most AI coding tools support some form of subagent delegation. The main agent spawns a helper, gives it a task, waits for the result. Simple in theory, painful in practice: ...

I wanted to see what happens when you give a team of AI agents a codebase and let them work on it with minimal supervision. Not a demo with three tasks, but a real, ongoing development project running for weeks. The result is ai-team – a multi-agent coordination system where specialized AI agents pick up tasks, write code in isolated git branches, get reviewed and tested by other agents, and merge their work back. As I write this, the system has processed over 1,000 task units (“beads”) and is currently building its own knowledge graph feature. ...

I wanted to understand how AI agents collaborate. Not in theory – I’ve read the papers – but in practice. What happens when you give a group of agents different roles, let them discuss, and have a human sit in as part of the team? So I built a system to find out. I call it agent-team, and it’s been my evenings-and-weekends project for the past couple of months. It’s a platform where AI agents with distinct roles – a coordinator, developers, testers, a lore master, whoever the project needs – work together on shared projects, communicating through persistent channels and threads, with a human participating as a team member rather than just an operator. ...