Review Agent Production Readiness

Use this skill as a second-pass production review for Agents SDK work. It is not a replacement for the official OpenAI Agents SDK skill or current OpenAI documentation.

Workflow

1. Establish the target workflow: user goal, input/output contract, tools, state, side effects, approval gates, runtime surface, and whether the design is single-agent, multi-agent, sandbox-backed, or explicitly orchestrated code.
2. If the request requires current SDK APIs, implementation changes, sandbox semantics, eval platform config, or deployment commands, read the official Agents SDK docs or use the official OpenAI Agents SDK skill first.
3. Read references/production-readiness-checklist.md for the applicable review lenses.
4. Review the real artifact whenever available: source code, prompts, tool schemas, eval cases, traces, logs, deployment config, runbooks, dashboards, or incident notes.
5. Report concrete gaps before general advice. Separate must-fix launch blockers from follow-up hardening.

Review Stance

• Prefer the simplest architecture that satisfies the workflow. Start with one agent and explicit code orchestration unless specialization, handoffs, or sandbox execution have a clear payoff.
• Treat every tool as a production interface: validate inputs, constrain permissions, bound runtime, make side effects idempotent where possible, and require approvals for irreversible or externally visible actions.
• Judge evals against behavior that matters: task success, required or forbidden tool calls, approvals, guardrails, state changes, traceability, failure handling, cost, and latency.
• Require observability that can diagnose real failures: trace IDs, tool-call logs, error classes, turn counts, latency/cost signals, escalation events, and enough context to reproduce issues without leaking secrets.
• Keep governance practical: scoped credentials, privacy controls, audit logs, rollout/rollback paths, human escalation, and ownership for incidents.