| name | agent-orchestration |
| category | cowork |
| description | Sequence specialist agents and humans around a single goal. Use for multi-step deliverables — e.g. research -> synthesis -> draft -> review -> publish. Triggers: "set up the pipeline", "orchestrate agents", "research-then-draft", multi-stage workflow. |
Agent Orchestration
You are the conductor. Each agent is one section of the orchestra.
Quick Reference (a pipeline has four parts)
| Part | What it is |
|---|
| Stages | Ordered list of agent / human steps |
| Gates | What must be true to move from stage N to N+1 |
| Handoffs | A packet between every two stages |
| Logs | Each stage's output captured for replay |
When to Use
- Multi-stage deliverables (research -> synthesis -> draft -> review -> publish).
- Any pipeline that crosses agent specializations or includes a human gate.
- When the same pipeline will run again (e.g., monthly competitive scan).
How to Use
- Map the pipeline end-to-end: which agent (or human) owns which step?
- Define gates between steps — what must be true to move on.
- Write the handoff packet for each transition (
agent-handoff).
- Insert human checks wherever a decision, sensitivity, or judgment call lives (
human-in-the-loop).
- Log the run in
memory/sessions/ so the orchestration can be replayed.
Example pipeline:
research-agent -> synthesis-agent -> HUMAN review -> comms-agent -> HUMAN ship gate
Examples
| Don't: Pipeline failure | Do: Pipeline success |
|---|
| Chain agents with no gates | Each transition has a gate condition stated |
| One human gate at the very end | Human gates wherever judgment / sensitivity / decision lives |
| Reuse yesterday's pipeline for a different shape | Re-design the pipeline for the new shape |
| Pipeline lives in your head | Pipeline written down in memory/sessions/<run>.md |
Critical Rules
- Gates between every two stages. No gate = poison flows downstream.
- Human-in-the-loop where judgment lives. Not at the end as an afterthought.
- Log every run. Replay is how the pipeline improves.
- Handoff packet at every transition (agent-handoff).
- Re-design per problem shape. Pipelines aren't reusable across problem types.
Common Pitfalls
- Chaining agents with no gates — first hallucination poisons everything downstream.
- Orchestrations no human ever inspects.
- Reusing yesterday's pipeline for a different shape of problem.
- Pipeline complexity that exceeds the deliverable's value.
- No log -> no replay -> no improvement.
Anti-Patterns
- Chaining agents without gates. The first hallucination poisons everything downstream.
- Orchestrations no human ever inspects.
- Reusing yesterday's pipeline for a different shape of problem.
- Over-orchestration — five agents where one would do.
Verify Before Running
Related
