The decisions-only orchestration discipline for AI agent fleets. One markdown skill, any model — Claude Code, Codex, Cursor, anything that reads markdown.
Agent
Panel judge — given 2+ independent proposals for the same problem, picks a winner and synthesizes. Use to close a panel run when one model's answer isn't trustworthy on its own. Proposers must never see each other's drafts; only the adjudicator sees all of them.
Agent
Cheapest-tier lookup agent — directory surveys, "where is X defined", git history, log/test-output summaries. Send it anything answerable by reading. It returns a synthesis, never file dumps. Use BEFORE the orchestrator reads more than 3 files to answer a question.
Agent
Mid-tier implementation agent — mechanical edits, well-specced features, test writing, first-pass reviews. Needs a complete brief up front (context, constraints, acceptance criteria, report format). Returns changes + a compact report, not a transcript.