I lead a team of agents, and I review every line.

1. Write it down before anyone writes code

Agents are only as good as what they're told. Every project starts with a written design and a plan broken into tasks small enough to verify one at a time, each with the exact test that proves it. This site was built that way: a spec, a sixteen-task plan, and a decision record that logs every judgment call an agent made on my behalf, with what it costs if it's wrong.

2. Make the gates do the arguing

I don't take an agent's "all tests pass" on faith. Work has to clear checks it can't talk its way past: tests written before the code, strict type checks, accessibility and contrast audits, and performance budgets. Some gates guard honesty, not just correctness. This site has tests that fail the build if any page claims a platform a project doesn't actually ship on.

3. Review it like a pull request

Every change gets read, and every feature gets driven for real: in a browser, and on a phone for anything that touches money. Screenshots get looked at, not just generated. That habit finds bugs no test was written for, like a tablet layout that quietly piled everything into one corner.

4. Build the tools when they're missing

When the workflow hits a wall, I build what it needs:

• Godot Pilot, a commercial bridge that lets an AI coding agent drive the Godot game editor directly.
• An MCP server that lets coding agents see and operate a desktop.
• A delegation layer that hands self-contained tasks from Claude Code to an open-weights model running locally on my own GPU, with every result reviewed and graded before it's accepted.
• A tool that measures and trims what an agent loads into context, to cut token costs.

5. Keep the decisions that matter human

Agents don't hold secrets, push releases, rewrite history, or publish anything on their own. Anything that's hard to undo or that faces the public waits for me. The speed comes from delegating the typing, not the judgment.

What this means on a team

I can set up this kind of workflow for a team, write specs that agents can actually execute, and review AI-written code with the same rigor as anyone's. If you're figuring out how your engineers should work with AI, that's a conversation I'd enjoy.