Project: Production Multi-Agent Pipeline
A production multi-agent system is not "five chatbots talking." It is a controlled workflow where specialized agents do bounded work.
Example product
Build a research-to-report pipeline:
text
planner -> researcher -> verifier -> writer -> reviewer -> human approval
Each agent has a narrow job, limited tools, and clear stopping rules.
Agent roles
| Agent | Job | Tools |
|---|---|---|
| planner | break task into steps | none or task database |
| researcher | gather evidence | search, retrieval |
| verifier | check claims | source lookup |
| writer | draft report | document editor |
| reviewer | find gaps and risks | eval rubric |
Do not share all tools with all agents
Tool access should be least privilege.
Bad:
text
every agent can browse, email, delete, deploy, and edit files
Better:
text
researcher can search
writer can draft
reviewer can comment
human approves external actions
State model
Track a shared task state:
- objective
- plan
- evidence collected
- claims
- open questions
- draft output
- review comments
- approvals
- final artifact
Agents should update state, not hide work in private chat history.
Failure handling
Add rules:
- max tool calls per agent
- max retries per failed step
- preserve partial work
- escalate after repeated uncertainty
- stop on conflicting evidence
- require approval before sending or publishing
Evaluation
Score the pipeline on:
- task completion
- source quality
- claim accuracy
- citation coverage
- unnecessary tool calls
- cost
- latency
- human edits required
Deployment path
text
offline traces -> internal users -> read-only tools -> limited write tools -> approvals -> gradual rollout
Knowledge check
Q1: Why should agents have narrow roles?
Narrow roles reduce confusion, unsafe tool use, and debugging difficulty.
Q2: What is the key production control for high-impact actions?
Human approval plus audit logging.