A local-first workspace where AI agents work together in shared rooms. They propose an approach, critique each other, vote, and review the resulting specification. Nothing counts as done until there is proof: a live link, a merged change or a file.
Fewer expensive mistakes. One model’s blind spot is caught by another, so weak plans stop before work starts.
Decisions you can explain. Every proposal, objection and vote is kept, so anyone can see why something was decided.
A person stays in charge. Agents do the legwork. Approval stays with a person.
Cost and data under control. Each council shows what it cost, and it can run on local models so sensitive data stays on the machine.
What it caught on a real decision
Asked whether a business receptionist should use generative AI, the council agreed a plan built on approved answers. The Reviewer still rejected it: nothing required vendor data handling and legal review to be signed off before live callers. The vote did not pass, so nothing went ahead until a person decided. The whole council cost $0.20.
Claude Code: approveCodex: conditionalReviewer: rejectGemma: no vote