16. Production Deployment Checklist

Use this checklist before considering an agent production-ready.

Task and Autonomy

• [ ] The task boundary is explicitly defined.
• [ ] Success criteria are measurable.
• [ ] Out-of-scope actions are listed.
• [ ] Autonomy level is chosen intentionally.
• [ ] Human approval points are defined for risky actions.

Context and Model

• [ ] System instruction defines role, constraints, and output protocol.
• [ ] Task description includes goal and success criteria.
• [ ] Context budget is defined before execution.
• [ ] Summarization or truncation strategy exists.
• [ ] Critical constraints are protected from loss.
• [ ] Model choice is justified for task, latency, cost, and reliability.

Structured Output

• [ ] Output protocol is explicit.
• [ ] Final-answer intent is required.
• [ ] Invalid output is rejected, not executed.
• [ ] Retry limit for malformed output is defined.
• [ ] Raw model output is logged separately from parsed intent.

Tools and Execution

• [ ] Tool set is minimal and task-relevant.
• [ ] Every tool has a clear description and argument contract.
• [ ] Tool allowlist is enforced.
• [ ] Argument validation is enforced.
• [ ] Permissions follow least privilege.
• [ ] Mutation tools are separated from read-only tools.
• [ ] Irreversible actions require approval or strong policy.
• [ ] Execution environment is sandboxed or otherwise bounded.
• [ ] Resource limits are enforced: time, memory, network, cost.

Loop Control

• [ ] Maximum iteration cap is enforced.
• [ ] Token budget is enforced.
• [ ] Cost budget is enforced.
• [ ] Timeout is enforced.
• [ ] Repeated-action detection exists.
• [ ] Stop reason is recorded for every run.
• [ ] Retry policy distinguishes transient errors from policy violations.

Memory and State

• [ ] Current state is explicit.
• [ ] Large artifacts are stored outside context and referenced by keys.
• [ ] Retrieved information is scoped and relevant.
• [ ] Stale state is superseded or expired.
• [ ] Sensitive state is access-controlled.

Safety and Observability

• [ ] Guardrails fail closed.
• [ ] Guardrail decisions are logged.
• [ ] Final answers are verified where possible.
• [ ] Every run has a stable run ID.
• [ ] Every step is traceable.
• [ ] Token, latency, and cost metrics are recorded.
• [ ] Failure paths are tested, not only happy paths.