How do you know an agent is ready for production?
An agent is ready when it has run the full job, end to end, against an environment that behaves like production, and handled the failures that environment throws at it. Passing unit tests and looking good in a demo are not the bar. Surviving state, permissions, timing, and real errors is.
The wrong question and the right one
The instinct is to ask "did the agent pass the tests?" For software that only computes an answer, that is enough. For an agent that acts on real systems, the right question is "did the agent survive the world it will operate in?" A demo shows the agent on its best day. Production is every other day.
A readiness checklist
- End to end, not step by step. The agent completed the whole workflow, not just isolated calls.
- State held up. What it created early was correct when read back late, with no duplicates on retries.
- Permissions were respected. When the environment rejected an action, the agent stopped.
- Failures were handled. A rate limit, rejected write, or late webhook did not derail the run.
- It was repeatable. The same scenario produced the same safe result more than once.
Why this is the layer teams skip
Recreating a realistic environment is harder than writing mocks, so most teams ship on the strength of layer one and two tests, then meet the real world and fail in ways nobody checked. Teams that avoid this run the agent against a sandbox that fails like production long before customers do.