An agent that passes testing got graded on its answer. Production grades its side effects.
That gap is the whole failure mode.
In a test run the world sits frozen. Fixtures don't move, every tool returns, and the only thing checked is the final output. In production the agent's plan is a sequence of writes into other people's systems, spread over minutes: a payout triggered, a status flipped, a document sent for signature. The run can die at step 5 with steps 1 through 4 already applied, and then the orchestrator retries from the top.
Three things I hold a production agent to.
Every write tool takes a caller-supplied idempotency key derived from the intent, not generated per attempt. Retries collapse instead of duplicating.
Preconditions get checked at the write, not at the decision. An agent that read a balance at step 1 and moves money at step 7 is acting on a number that has aged.
Reads and writes are split at the tool boundary. The agent explores freely, and the dangerous surface stays small enough to audit.
Evaluation follows from that. Scoring final answers tells you the model is fine. Replaying traces tells you what state the world was left in when a run stopped halfway.
If you run agents against real systems, I'd start with the thing that happens today when yours stops at step 5.