Open-source Triad protocol splits AI coding agents into fixed roles
Triad is an protocol built to fix a problem that shows up fast when multiple work together: if every agent can plan, edit, review, and commit, the roles blur and review becomes hard to trust. Its default setup assigns Codex as a read-only architect and reviewer, Claude Code as the primary implementer that writes the actual code, Qwen Code as an optional junior for very basic tasks, and a human as the final authority. These roles aren't fixed to specific tools — Triad is adapter-driven, so any other CLI tool can take over a role without changing the underlying engine.
For each project, Triad keeps an auditable `.pair/` directory holding the , plan, session IDs, review verdicts, suggestions, , and a full exchange log. The implementer works on one task at a time, and the architect reviews the uncommitted diff against the original before anything is accepted. The junior role is deliberately constrained: every task needs separate human approval, one approval allows exactly one attempted run, and failed junior work goes straight back to the primary implementer instead of retrying automatically.
Long working sessions are treated as disposable memory, with anything that needs to persist stored separately.
Key points
- Four-role setup: Codex as read-only architect/reviewer, Claude Code as implementer, Qwen Code as optional junior, human as final authority
- Auditable `.pair/` directory logs , plan, review verdicts, and the full exchange history
- Implementer handles one task at a time; architect reviews the uncommitted diff against
- Junior tasks require separate human approval each time, and failures return to the implementer with no auto-retry
- Adapter-driven design lets other CLI tools fill any role without changing the core engine