OpenAI Agents platform

Agents act in outside systems through function tools the developer defines, which pause the turn until the application returns a result, and through MCP servers, either remote ones reachable from OpenAI or servers started inside the session's own environment over HTTP or stdio, with tools discovered and called by the platform. Hosted tools add web search, file search, a shell and a code interpreter, so one agent can look something up, read company files and run code in the same turn.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/tools/mcpread 2026-11-09

In the Agents API, in beta, OpenAI runs the orchestration: a session is a durable unit of work made of turns, and the platform handles compaction and recovery across long tasks. A main agent can delegate independent tasks to subagents, each with its own context, that work in parallel while the main agent coordinates and combines their results, such as comparing several release notes at once. Teams building in code get the same multi agent patterns, handoffs between agents and agents used as tools, in the open source Agents SDK.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/multi-agentread 2026-11-09

File search lets an agent search a company's own files for relevant passages before it answers. Documents are uploaded once into vector stores that persist and are billed for storage by the gigabyte per day, with the first gigabyte free, so the index outlives any single run and grows past what fits in a prompt. Embeddings are also available directly for teams that run their own retrieval, and files placed in a session's environment give an agent working material for a task.

Sourcedevelopers.openai.com/api/docs/guides/tools-file-searchread 2026-11-09

Agents in the Agents API, in beta, raise approval requests that wait as required actions until the application answers them, so a person can approve or reject a step before it runs. Automatic review can approve or deny supported requests by policy for the main agent and every subagent, can change between tasks, and never overrides the environment's network restrictions. Browser tasks ask before visiting each new website and hand sign in back to a person, and the Agents SDK adds guardrails and human review for agents built in code.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/sessions/approvalsread 2026-11-09

Role based access governs both the dashboard and the API: organization roles reach every project, project roles stay inside one, custom roles carry exactly the permissions a team chooses, and groups, project service accounts and Admin API keys round out who can do what. Organizations and projects can restrict API calls to allowlisted IP ranges or Azure Private Link. OpenAI's SOC 2 Type 2 report and its ISO 27001, 27017, 27018 and 27701 certificates cover the API platform, alongside ISO 42001 and CSA STAR.

Sourcedevelopers.openai.com/api/docs/guides/rbacread 2026-11-09

Tracing is on by default for Agents API sessions, in beta. Each session lists its status, model, turns and token usage, and each turn's trace shows the model responses, tool calls and work delegated to other agents in order, so a team can see why an agent did what it did. Traces can be read in the dashboard or exported through the API as OTLP JSON into the team's own observability stack, and agents built with the Agents SDK record model calls, tool calls, guardrails and handoffs the same way.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/tracingread 2026-11-09

Sessions in the Agents SDK carry an agent's conversation memory across separate runs, stored and read back automatically under a session ID in a backend the team chooses: SQLite, Redis, SQLAlchemy, Dapr, MongoDB, encrypted storage or OpenAI's Conversations API. Memory operations let the application read, correct or clear one session's history without touching other data. In the Agents API, sessions stay durable across turns with automatic compaction, and saved work can be deleted when a task ends.

Sourceopenai.github.io/openai-agents-python/sessionsread 2026-11-09

Data residency is a project setting: a team selects the region where customer content is stored at rest, and in regions that support it, where inference runs, after confirming eligibility with sales, with a 10 percent uplift on eligible newer models. Agent work can run in an OpenAI hosted sandbox or on the customer's own environment provider, E2B, Daytona, Blaxel, Cloudflare, AWS Lambda MicroVMs, Databricks or DigitalOcean, and approved customers can switch on zero data retention.

Sourcedevelopers.openai.com/api/docs/guides/your-dataread 2026-11-09

Ready made material comes as parts to assemble rather than agents to switch on. Complete examples show an agent writing and running a script in a hosted sandbox, comparing release notes with subagents and working with policy skills and specialist agents, ChatKit ships widgets and actions for building an agent's chat interface, and the cookbook covers agents, evals and multimodal patterns. Agent Builder's starting templates go away when that product closes, and its workflows move to the Agents API.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/overviewread 2026-11-09

Scheduled automations in the Agents API, in beta, start a new session or continue an existing one from a cron expression and time zone, with no request from the application, and an agent given the automations tool can schedule its own follow up work in the middle of a task. Session webhooks report lifecycle changes so the developer's systems react when a session needs input or finishes, and applications can also start sessions from any backend event through the API.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/sessions/automationsread 2026-11-09

Agents on the Agents API and the hosted tools run on OpenAI's own models, with the team choosing the model per agent and paying that model's API rate, and Amazon Bedrock Managed Agents runs the same API on AWS rather than adding other model makers. The open source Agents SDK is the route to other providers: it can mix models across makers in one workflow through Any-LLM and LiteLLM adapters, so choice across makers reaches agents built in code but not the managed service.

Sourceopenai.github.io/openai-agents-python/modelsread 2026-11-09

Everything on the platform is an API first. The REST reference covers agents, sessions, turns, environments and automations alongside the Responses API, official client libraries wrap them, the open source Agents SDK in Python and TypeScript builds agents in code, and ChatKit offers a frontend kit plus an advanced mode that connects to the team's own server. Admin APIs and Terraform resources manage projects, roles, service accounts, rate limits and spend.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/overviewread 2026-11-09

OpenAI Red Teaming is OpenAI's own offer for new evaluation work: a managed service, sold to enterprise customers, that probes a team's AI applications and agents with adversarial, abusive and unexpected inputs before deployment and supports coordination, review and reporting on the findings. Promptfoo, the open source eval and red teaming framework that OpenAI owns, is a separate product sold under its own name with its own enterprise plans. The Evals dashboard and API, with their graders, trace grading and prompt optimizer, are read only and shutting down, so they take no new work.

Sourcedevelopers.openai.com/api/docs/guides/red-teamingread 2026-11-09

Agents navigate websites and work through browser interfaces in a browser OpenAI hosts: the application starts a browser session, sends a task and follows events while the agent decides each next step from what it sees, asking for access to each new website and handing sign in to a person when an account is needed. The computer use tool also runs browser or desktop tasks through the SDKs, or connects to the customer's own runner for its own machines.

Sourcedevelopers.openai.com/api/docs/guides/agents-api/tools/computer-useread 2026-11-09