OpenAI Agents SDK Production Playbook

Look forPin the package, because minor 0.Y.Z releases may break public interfaces. Set the model explicitly or through OPENAI_DEFAULT_MODEL or RunConfig: with none set, agents run gpt-5.6-luna with reasoning effort none and low verbosity.

Look forTracing is on by default and sends runs to OpenAI, with model and tool inputs and outputs included. add_trace_processor keeps OpenAI's exporter, and only set_trace_processors replaces it, as one builder on X warns. Organizations under Zero Data Retention cannot use the built-in tracing, and OpenAI's tracing and data-control pages state no retention period for traces.

Look forBuilders report traces silently missing from long-running workers and whole batches dropped on HTTP 429. Call flush_traces after each trace in background workers, and check that a run in each worker type reaches the dashboard or your own backend.

Look forSQLite sessions suit one process. One builder wrote a Redis store before OpenAI shipped RedisSession, and an open issue reports RedisSession failing on Redis Cluster. Pick SQLAlchemy, Redis, MongoDB, Dapr, or OpenAI Conversations sessions, and test resume after a restart.

Look forMark data-changing tools as needing approval and store RunState while a run waits. Guardrails run in parallel by default: one builder measured all 6 tool calls started before the tripwire fired in 30 of 30 runs. Use blocking guardrails where that matters, and bind approvals to their run, since an open issue shows one approval authorizing a call in another run. On 0.22.0 an approval resume could leave an orphaned tool output that failed every later turn, so test resume on the version you pin.

Look forgpt-5.6-luna is not supported on the Free tier and starts at 5,000 RPM and 2,000,000 TPM on Build. OpenAI's status page logged elevated gpt-5.6-luna API errors on October 21, 2026 and Responses API errors on October 22, so plan retries and a fallback model.

Look forMeasure a run of real length. Builders report agent calls that slowed from under 4 to over 11 seconds on GPT-4.1, nested agents as tools that never hit the prompt cache, and long tool loops that grew slower each turn. Set spend alerts and a hard spend limit at the organization or project level, since alerts alone never stop traffic.

Look forOther models run through the LiteLLM and Any-LLM adapters, which OpenAI calls best-effort beta, or the AI SDK adapter in TypeScript. Builders report a Gemini crash on cached tokens, truncated responses returned as success, and parallel tool calls failing on Gemini, so test each model end to end.