Multi-Agent Software Team
Run planner, coder, and reviewer agents over a shared repo dict until fake tests pass — with budgets, patches, and no infinite rewrite loops.
Outcome: A three-agent coding team that plans, edits a dict-as-repo, reviews, and stops when a fake test runner is green — or when the budget dies.
advanced · 8–12 hours
- 01
Overview and Architecture
Split planning, coding, and review into roles that share a repo dict and a test oracle, and define stop as tests-pass — not as 'the coder felt done'.
- 02
Shared Repo and Test Oracle
Implement the repo dict, path allowlists, full-file writes, and a fake test runner that execs fizzbuzz in an isolated namespace.
- 03
Planner, Coder, and Reviewer Policies
Implement three fake role policies with narrow outputs: a plan, a source patch, and a structured review — then take turns in code.
- 04
Supervisor Loop Until Tests Pass
Turn-take planner, coder, tests, reviewer; apply allowlisted patches; stop on green tests or max rounds.
- 05
Evals and Hardening the Team
Test forbidden paths, budget stops, two-round success, and block dangerous source before exec.