Agent Production Readiness Checklist (2026 Edition)

Use this checklist to move from a working demo to a production agent that stays reliable, stays within budget, and leaves an audit trail.

1) Define the workflow boundary
- Name one workflow (examples: “triage support tickets,” “enrich inbound leads”).
- List the systems the agent may touch (Zendesk, Jira, Salesforce, internal DB).
- Write the success condition for a single run (what must be true at the end).

2) Write an authority spec (permissions)
- Split read-only tools from write tools.
- Flag high-risk actions (refunds, cancellations, deletions, exposing PII).
- Specify approval rules (required human review, dual control, or auto-approved).

3) Build a tool gateway (avoid direct SaaS calls from the agent)
- Implement per-tool and per-method allowlists.
- Enforce tenant isolation and least privilege.
- Use short-lived scoped tokens; never place secrets in prompts.

4) Enforce budgets in code
- Step budget: cap tool calls per run.
- Spend budget: cap tokens and cost per run.
- Rate limits: per tool and per tenant; include backoff and bounded retries.

5) Implement trace logging and retention
- Log the user request, prompt/version, tool calls, tool responses, policy decisions, and final output.
- Store traces immutably with a retention policy aligned to your compliance needs.

6) Build an evaluation set from real cases
- Collect historical examples for the workflow.
- Label outcomes (success, partial success, failure) and tag by risk.
- Add edge cases (timeouts, missing fields, ambiguous requests).

7) Use metrics that map to operations
- Task success rate (segmented by risk tier).
- Cost per successful task (including retries and tool calls).
- Latency (median and tail).
- Escalation rate with a clear “reason for escalation” taxonomy.

8) Release in phases
- Phase A: shadow mode (agent runs; humans execute) with trace review.
- Phase B: human-in-the-loop (agent drafts; humans approve).
- Phase C: limited autonomy for low-risk actions with tight monitoring.

9) Add safety rails for write actions
- Require structured outputs and schema validation.
- Use explicit confirmations for irreversible changes.
- Add “undo” paths where possible (revert field changes, cancel queued actions).

10) Make ownership real
- Assign an on-call owner and escalation path.
- Maintain a rollback plan (model revert, prompt revert, tool kill switch).
- Review traces on a schedule; refresh eval sets as the workflow changes.

If you can’t answer “what can this agent do, how much can it spend, and how do we audit its actions?” it’s not production-ready.