Agents do the work. People decide.
We are a small team, and most of the work at Codeclowns is done by AI agents. That only works because every decision is made by a person and written down, so the next agent starts from it instead of from zero.
- 1 · peopleIdeaA problem worth solving, written down in the product hub.
- 2 · agentsResearchParallel agent lanes: market, users, feasibility, competitors. Every claim cites a source.
- 3 · agentsDebateA sceptic and a builder argue it out, so the weak points surface before we commit.
- 4 · peopleDecideA person writes and signs the decision record. Agents never decide alone.
- 5 · agents + peopleBuildCoding agents write the code; people review every change.
- 6 · agents + peopleVerifyThe result is checked against the decision before it ships.
- 7 · peopleLearnWhat we found, including what we got wrong, goes back into the hub.
Why it works
Agents are fast and tireless, and confidently wrong often enough that nothing they produce should go straight into a product. So the loop puts a person at the two points where judgement matters most: deciding what to build, and checking that what was built matches the decision.
Research runs as several agents in parallel, each with one question, a list of sources it must cite, and a fixed format for its answer, so the results merge cleanly. Before anything is decided, two agents argue opposite sides: one builds the case, the other tries to break it. The decision record that a person signs names the options that lost and why.
Coding agents then build against that record. Every change is reviewed by a person, and before it ships it is checked against the decision it was meant to carry out. What we learned, including what we got wrong, goes back into the product hub, where it is the starting point for the next piece of work.