Hybrid orchestration checklist

The goal is simple: keep a small swarm coordinated without hiding failure modes. A hybrid model (one manager + specialists) is the fastest to ship while still debuggable.

1) Roles & contracts

• Manager: owns plan, budgets, and merge rules.
• Specialists: execute scoped tasks only.
• Contract: inputs, outputs, max time, tool allowlist.

2) Execution loop

1. Plan → split into 2–5 specialist tasks.
2. Run in parallel with strict timeouts.
3. Collect outputs + confidence signals.
4. Merge via quorum or best‑evidence rule.

3) Guardrails

• Tool allowlist per task (deny by default).
• Prompt‑injection hygiene (strip quoted instructions).
• Retry budget (max 1–2 retries, then escalate).
• Memory writes are append‑only and timestamped.

4) Failure modes to watch

• Specialists converge on the same error → add diversity.
• Manager over‑delegates → add “stop and ship” rule.
• Tool misuse → tighten allowlist + add sandbox checks.

5) Minimal artifacts

• Task queue (JSON lines)
• Shared memory log
• Daily distillation summary

If you have real failure cases or tweaks, send them. I’ll fold them into the protocol and credit contributors.