# Agent reliability checklist

Use this before adding more prompts, tools or model capacity to an AI workflow.
It is a practical starting point, not a guarantee that a workflow is safe for
production.

## 1. Define the loop

- [ ] What input starts the workflow?
- [ ] What observable output counts as useful?
- [ ] Which decisions remain with a person?
- [ ] What data is allowed to leave the system?
- [ ] What is the stop condition when the workflow cannot proceed?

## 2. Make failures explicit

- [ ] Timeouts and retry limits are defined for every external call.
- [ ] Retries use backoff and do not duplicate side effects.
- [ ] Writes are idempotent or have a reconciliation path.
- [ ] Invalid or incomplete model output is rejected before it reaches a user.
- [ ] A human checkpoint exists for high-impact or irreversible actions.

## 3. Measure before optimizing

- [ ] Record latency, error rate, retry count and cost per run.
- [ ] Keep a small evaluation set that represents real edge cases.
- [ ] Compare changes against the same inputs and acceptance criteria.
- [ ] Record the model, prompt and tool versions used for each evaluation.
- [ ] Define a rollback path before changing the production loop.

## 4. Operate the workflow

- [ ] Logs identify a run without storing secrets or unnecessary personal data.
- [ ] Alerts distinguish an outage from a degraded model response.
- [ ] Permissions limit what the agent can read and change.
- [ ] A person can pause the workflow without deleting its audit trail.
- [ ] The handoff explains what happened, what remains and who owns the next step.

## A useful next step

Pick one real workflow and write down its current latency, failure modes and
manual touchpoints. If those constraints are still unclear, an audit should
start with measurement rather than a larger build.