F13: Coordinator workflow patterns (supervisor / debate / pipeline / group-chat)

[frisbeeman]

Part of the v0.4 — Multi-Agent Orchestration roadmap.

Summary

Library of named coordinator recipes, each as a typed wrapper builder + a runnable example:

• Supervisor — one parent dispatches subtasks to workers, aggregates results. (Child workflows.)
• Debate / vote — N agents independently answer; a judge agent picks the winner. (Parallel child workflows + final selection.)
• Pipeline — sequential handoff: A's output is B's input. (Sequential child workflows.)
• Group chat — N agents share a mailbox, round-robin or moderator-driven. (#F12 + a moderator workflow.)

Why

These are the canonical multi-agent patterns. Shipping them as named, typed recipes (vs. having users build from primitives every time) is the differentiator vs. non-durable frameworks.

Scope

• New src/coordinators/ module — one file per pattern.
• Each pattern: typed builder + runnable example.
• Reference ""research crew"" example: orchestrator + 3 worker agents (search, summarize, fact-check) + final synthesis.
• Per-pattern documentation of edge cases (cycles, deadlocks, partial failures).

Files

• new src/coordinators/ module (4 files)
• 4 examples

Depends on

• #F11 (child workflows)
• #F12 (mailbox) — for group-chat

Effort & risk

• Effort: Large (~600 LOC across 4 patterns + 4 examples).
• Risk: Medium — each pattern's edge cases need documenting.

Acceptance

• Each coordinator has a runnable example.
• Each pattern has a benchmark showing throughput characteristics under failure.
• Reference research-crew example end-to-end.