Note
AI agents in my own stack
The agent setup I run every day. Several coding agents, isolated worktrees, one shared memory. My own stack is the laboratory.
AI Agents Systems
Before I recommend a system to anyone, it has to survive my own daily use. This note tracks the agent setup I work with right now. It changes often, and that’s the point.
Several agents, side by side
I run several coding-agent CLIs in parallel. Sometimes I point two of them at the same task and compare the actual work, not the plans. Different harnesses have different strengths, and the only way to find out is to make them do the same job.
One worktree per agent
Every agent session gets its own isolated git worktree. Parallel sessions can’t trample each other’s work, and a failed experiment is cheap to throw away. I learned this the hard way when two sessions sharing one checkout landed commits on each other’s branches.
A shared memory
All my agents read the same file-based knowledge base at session start. Plain markdown, versioned in git, one note per durable fact. I wrote up how it works and why it matters in Giving my agents a shared memory.
Always on
An always-on machine runs headless agents I can reach from my phone. Tools connect over MCP, the same protocol across every harness, so the parts stay swappable.
Why this is public
This stack is the laboratory behind coaching founders on AI systems. What survives here is what I bring to others.