How do you observe a multi-agent workflow?
Why C is the answer
Use logs, artifacts and operational signals from each agent
Logs, artifacts and operational signals show how the work happened. A final result alone hides that.
In detail
- Logs, artifacts and operational signals show each agent's work.
- Collect signals in one place.
Use case
An illustrative scenario, not a real incident.
A workflow slows down. Logs and signals from each agent show that one tool is timing out, which a final pull request could never reveal.
See it in different ways
sequenceDiagram participant W as Workflow participant A as Agent participant H as Human W->>A: Step A--xW: Stalls or fails review W->>W: Pause later steps W->>A: Retry or roll back W->>H: Escalate if it still fails
mindmap
root(("Observability"))
n1["Logs"]
n2["What each agent did"]
n3["Artifacts"]
n4["What it produced"]
n5["Signals"]
n6["Timing and failures"]Key point: Observe each agent, not just the result.
Why the other options are wrong
- A. Watch the screen of the machine that runs each agent in real time
- Watching screens does not scale, and it leaves no record.
- B. Ask each of the agents for a summary after the workflow is finished
- Agent summaries may leave things out.
- D. Read the final pull request only, because it contains everything
- A final pull request hides the handoffs.