Skip to main content

Decision brief

OpenAI Agents SDK Review: A Small Agent Runtime Built Around OpenAI

Who the OpenAI Agents SDK fits, from Harvey's evals to teams leaving Agent Builder before Nov 30, 2026, what its default model and tracing send to OpenAI, and the hosting, state, and version decisions to settle first.

Reading time
6 min read
Sources
30 cited
In this piece
OpenAI Agents SDK−17 Mixed

GILD verdict

Choose the OpenAI Agents SDK when your agents run on OpenAI models, need handoffs, guardrails, and approvals, and can live inside a Python or TypeScript service your team hosts.

The workflow
A team building agents on OpenAI models into its own service, where agents hand work to each other, call tools and MCP servers, and pause for a person's approval.19
The decisive evidence
  • DocumentedThe SDK runs agents with tools, handoffs that the model sees as transfer tools, agents as tools, guardrails, sessions, and approvals that serialize a paused run, with tracing on by default.189
  • DocumentedWith no model set, agents run gpt-5.6-luna with reasoning effort none and low verbosity. Other models go through the LiteLLM or Any-LLM adapters, which OpenAI calls best-effort beta, or through the AI SDK adapter in TypeScript.234
  • DocumentedTraces go to OpenAI's backend unless set_trace_processors replaces the exporter, they include model and tool inputs and outputs by default, and organizations under Zero Data Retention cannot use them.5
  • GILD analysisOf the 59 public accounts on this product's Signals page, 11 praise the small set of primitives and 9 report agents built on it, from Lark's sandbox runtime to Harvey's evals on the SDK's traces. Runtime defects draw concerns in 14 of them, latency in 8, and the OpenAI-first design in 8, most of them closed GitHub issues. The record leaves open how the SDK holds up under sustained company load, since no named team publishes request volume or uptime on it.
The tradeoff
You trade portability for a small runtime shaped around OpenAI's platform. Tracing, hosted tools, and Responses-only features work best on OpenAI models, and the SDK stays on 0.Y.Z versions in which a minor release may break public interfaces.2510
What would change the call
The call changes if the agent must run models from many vendors on one code path or needs explicit, checkpointed control flow. Then LangGraph or the Vercel AI SDK fits better, and OpenAI's hosted Agents API suits teams that would rather not host the runtime. Teams on Agent Builder have until November 30, 2026, when OpenAI retires it and points code workflows to the SDK.22022

OpenAI Agents SDK

Agent Framework by OpenAI

−17 Mixed
Product page

Fit and limits

Who it fits

  • The agents run on OpenAI models and the team wants handoffs between specialist agents without writing a graph.
  • Tools that change data need a person's approval, and a paused run must survive until the decision arrives.
  • The team can host a Python 3.10+ or TypeScript service and choose its own session store.
  • Traces can go to OpenAI's dashboard, or the team will route them to its own backend with a custom processor.
  • Evals must run on real agent traces: Harvey exports the SDK's OpenTelemetry traces to LangSmith for offline evaluation of its legal search agent (Harvey blog, Dec 10, 2025).
  • The team has workflows in Agent Builder, which leaves the OpenAI platform on November 30, 2026; OpenAI recommends the SDK for workflows that continue as code.

Where it falls short

  • Organizations under Zero Data Retention, or with prompts that must stay in-house: built-in tracing sends model and tool inputs and outputs to OpenAI by default and is unavailable under Zero Data Retention, so traces need another processor and backend.
  • Teams that need models from many vendors: the LiteLLM and Any-LLM adapters are best-effort beta, and Responses-only features are rejected on other backends.
  • Agents that must not act before a check passes: guardrails run in parallel with the agent by default, so tools can start before a tripwire fires unless blocking mode is set.
  • Teams that cannot absorb breaking changes: the SDK is pre-1.0, and a minor version may change public interfaces that are not marked beta. On 0.22.0 an approval resume could leave a tool output with no tool call in the session, which failed every later turn of that conversation until a fix landed.
  • Designs built around async code and streaming: Aurelio Labs used the SDK extensively and left it, with LangChain and LangGraph, for an async framework of its own in October 2025.

Implementation reality

  1. Pin the version and the model

    Pin openai-agents or @openai/agents, and set the model on each agent or through OPENAI_DEFAULT_MODEL so you don't rely on the default, which is gpt-5.6-luna today. Read the release notes before each minor upgrade, since minor versions may break public interfaces.

  2. Decide where traces go

    Tracing is on by default and exports to OpenAI. Use set_trace_processors to replace the exporter with your own backend, turn off trace_include_sensitive_data if prompts hold customer data, and call flush_traces in long-running workers.

  3. Move sessions to shared storage

    Use a Redis, SQLAlchemy, MongoDB, or Dapr session for agents that run on more than one instance, and test resume after a container restart.

  4. Gate tools and test guardrails

    Mark tools that change data as needing approval, store RunState while a run waits, and use blocking guardrails where a tool must not start before the check passes.

What to verify before you buy

Models and limits

  • gpt-5.6-luna limits: not supported on the Free usage tier. The Build tier starts at 5,000 RPM and 2,000,000 TPM, and tiers rise automatically with spend.
  • Responses-only features: rejected on Chat Completions models and non-Responses backends, so test each model you plan to run.

Access

  • Single Sign On: OpenAI's trust portal says Single Sign On and domain verification enable access controls for select OpenAI products.

Data retention

  • Abuse monitoring logs: by default, OpenAI generates them for all API feature usage and keeps them for up to 30 days, unless longer retention is required by law or is reasonably necessary to protect its services or any third party from harm.
  • Zero Data Retention: eligible customers can have their content excluded from these logs by getting approved for Zero Data Retention or Modified Abuse Monitoring, which currently need OpenAI's prior approval and acceptance of additional requirements, subject to limitations in its data controls guide.

What to measure in your pilot

  • How the SDK holds up under sustained company load: Harvey describes its evals on the SDK's traces, and no named team publishes request volume or uptime.
  • How many teams move from Agent Builder to the SDK before Agent Builder leaves the OpenAI platform on November 30, 2026.
  • Whether the open Redis Cluster and approval-reuse issues are fixed in the next releases.
  • How OpenAI's hosted Agents API, in public beta since September 10, 2026, changes the case for hosting the SDK yourself.

Sources

Official documentation
28
Public source
2
  1. Agents SDK docs (Python): Overview (primitives and the Responses API default)Official documentationopenai.github.io
  2. Agents SDK docs (Python): Models (default model, OPENAI_DEFAULT_MODEL, LiteLLM and Any-LLM)Official documentationopenai.github.io
  3. Agents SDK docs (TypeScript): ModelsOfficial documentationopenai.github.io
  4. Agents SDK docs (TypeScript): AI SDK extension for other modelsOfficial documentationopenai.github.io
  5. Agents SDK docs (Python): Tracing (default exporter, processors, Zero Data Retention)Official documentationopenai.github.io
  6. Agents SDK docs (Python): SessionsOfficial documentationopenai.github.io
  7. Agents SDK docs (Python): Guardrails (parallel and blocking execution)Official documentationopenai.github.io
  8. Agents SDK docs (Python): HandoffsOfficial documentationopenai.github.io
  9. Agents SDK docs (Python): Human in the loop (approvals and RunState)Official documentationopenai.github.io
  10. Agents SDK docs (Python): Release process (0.Y.Z versioning)Official documentationopenai.github.io
  11. Agents SDK docs (TypeScript): TracingOfficial documentationopenai.github.io
  12. OpenAI API docs: Agents integrations and observabilityOfficial documentationdevelopers.openai.com
  13. OpenAI API docs: gpt-5.6-luna (price, context window, rate limits by usage tier)Official documentationdevelopers.openai.com
  14. OpenAI API docs: Pricing (per-token prices)Official documentationdevelopers.openai.com
  15. OpenAI Help Center: Troubleshooting API usage and spend limitsOfficial documentationhelp.openai.com
  16. OpenAI API docs: Rate limits and usage tiersOfficial documentationdevelopers.openai.com
  17. OpenAI API docs: Pricing (tools and containers)Official documentationplatform.openai.com
  18. OpenAI API docs: Data controls in the OpenAI platform (Zero Data Retention, store behavior)Official documentationdevelopers.openai.com
  19. OpenAI: Enterprise privacy (training, retention, SAML SSO, SOC 2 Type 2)Official documentationopenai.com
  20. OpenAI: Introducing the Agents API (Sep 10, 2026)Official documentationopenai.com
  21. OpenAI API docs: Agents API overviewOfficial documentationdevelopers.openai.com
  22. OpenAI: Introducing AgentKit, with the Jun 3, 2026 update retiring Agent Builder on Nov 30, 2026Official documentationopenai.com
  23. OpenAI: The next evolution of the Agents SDK (Apr 15, 2026)Official documentationopenai.com
  24. Harvey blog: How agentic search unlocks legal research intelligence at Harvey (Dec 10, 2025)Public sourceharvey.ai
  25. Agents SDK Python issue #4827: an approval resume on main after #4613 leaves an orphaned tool output (Sep 2, 2026)Public sourcegithub.com
  26. openai-agents-python on GitHub (MIT license)Official documentationgithub.com
  27. openai-agents-js on GitHub (MIT license)Official documentationgithub.com
  28. openai-agents on PyPI (release history)Official documentationpypi.org
  29. @openai/agents on npmOfficial documentationnpmjs.com
  30. OpenAI Trust PortalOfficial documentationtrust.openai.com

Keep reading

  1. Buyer guideOpenAI Agents SDK Production PlaybookPin the version and the model, decide where traces go, move sessions to shared storage, then test guardrails, approvals, and cost before users arrive.
  2. ResearchLangGraph Adoption, Releases & Support Window (2026)LangGraph's downloads, GitHub activity, and release pace next to CrewAI and OpenAI's Agents SDK, with LangChain's 1.x long-term-support commitments and how well it has kept them.
  3. ResearchOpenAI Agents SDK in Practice: What Builders ReportIn 64 public accounts, from Harvey's evals on the SDK's traces and Lark's sandbox runtime to solo developers, builders report what its small set of primitives does well and where tracing, latency, state, and other models cause trouble.
Read the research