An Agent Workflow Checklist

Questions before you publish

Are inputs and state variables explicit and typed, with identity from your backend? Does each agent have one job, a suitable model, few tools and a structured output where others depend on it? Do branches test structured fields with an else path? Do loops have caps? Do risky actions require approval with reject and timeout paths? Are guardrails tested for both attacks and false positives? Do all paths end in one output contract? Has the evaluation set passed, with routing measured? Is the version described, diffable and rollback-ready? Are cost, latency and traces monitored?

The checklist

Use it at every publish.

[ ] typed inputs + state; identity from backend
[ ] one job per agent; structured outputs where consumed
[ ] branches on fields; explicit else
[ ] loops capped; cap path handled
[ ] approvals for write/send/pay/delete; reject + timeout paths
[ ] guardrails tested (attacks + normal traffic)
[ ] least-privilege tools and MCP servers
[ ] single output contract on every path
[ ] eval set passed; routing accuracy tracked
[ ] version notes, diff reviewed, rollback ready; cost + traces monitored

Re-check after model changes

When you switch a node's model, rerun the evaluation set; routing and tool use can shift.

Quick check: Which item belongs on a pre-publish checklist?

• Every loop has an iteration cap and the cap path is handled
• Give every agent every tool
• Leave else branches empty
• Skip evaluation for small changes