Part 7: How the Agent Learns From Its Mistakes
Patching a hole in my agent's review process pulled a thread: the fix needed memory, the memory got too specific, and cleaning that up taught the system where knowledge belongs....
Read More →Tag: ai-agents
From Coding Agent to Chief of Staff
I'd been using Hermes to write code. Then I pointed it at my house: one always-on agent for mail, meals, trips, and bills -- and the guardrails that let us...
Read More →Part 6: The Autonomy Ladder
Every external action runs through an autonomy gate with three settings -- gated, push-draft, and full -- so the coordinator does as much as I trust it to and no...
Read More →Part 5: GitHub Issues as the Agent's Backlog
I made GitHub Issues the agent's backlog -- the coordinator creates structured issues from raw ideas, triages overnight, and grooms the backlog weekly, all behind an opt-in gate.
Read More →Part 4: Running Untrusted Tools Safely
The moment you let an agent inject code it didn't write, you need a system that answers one question: is this safe to run? Four scripts, one pipeline, zero trust...
Read More →Part 3: Dynamic Tool Discovery and Injection
The coordinator now discovers and injects task-specific skills from trusted external sources at dispatch time -- so the coding engine gets exactly the domain knowledge it needs, not a blob...
Read More →Part 2: One Coordinator, Swappable Coding Engines
I pulled the coding engine out into a swappable layer so the coordinator can dispatch to Antigravity, Claude Code, OpenCode, or a local model running on my machine — same...
Read More →Part 1: I Built an Always-On AI Coding Agent That Plans, Codes, and Reviews Its Own Work
I took a 2019 MacBook Pro that was collecting dust and turned it into an autonomous coding coordinator. It plans before it codes, delegates to a separate AI for implementation,...
Read More →