OpenAI Agents SDK (Python)

What it does

Writes code

YesSandbox agents get a workspace with apply_patch file editing and shell, documented for coding tasks; the feature is in beta.Sandbox Agents in the Agents SDK give the model a persistent workspace where it can search large document sets, edit files, run commands, generate…openai.github.io/openai-agents-python/s… ↗

Debugs and fixes

YesThe documented sandbox use case is orchestrating automated fixes for GitHub issue reports and running targeted tests.coding and debugging, for example orchestrating automated fixes for issue reports in a GitHub repo and running targeted testsopenai.github.io/openai-agents-python/s… ↗

Writes tests

PartlyAdding a regression test is given as a sandbox multi-step workflow example; no dedicated test-generation feature is documented. The SDK's own testing utilities test agent workflows, not application code.multi-step workspace tasks, for example fixing a bug in one run and adding a regression test later, or resuming from snapshot or sandbox session stateopenai.github.io/openai-agents-python/s… ↗

Reviews code

PartlyFile-grounded review and per-reviewer sandbox workspaces are documented use cases; no diff or pull-request review surface is described.file-grounded review or analysis, for example checking onboarding packets, generated reports, or artifact bundles before answeringopenai.github.io/openai-agents-python/s… ↗

Whole-codebase work

PartlyA manifest can clone a Git repository into the sandbox workspace and snapshots resume prior work, but no repository-wide migration capability is claimed.You define the workspace around the data the agent needs. It can start from GitHub repos, local files and directories, synthetic task files, remote…openai.github.io/openai-agents-python/s… ↗

Not established: Ships it

How much it does unattended

Runs autonomously

YesAgents run a built-in loop until the task completes; human approval is opt-in per tool.Agents: Build agents with instructions, tools, guardrails, handoffs, and a built-in loop that continues until the task is complete.openai.github.io/openai-agents-python ↗

Multi-agent

YesHandoffs switch the active agent; Agent.as_tool() lets a manager agent call specialists. An experimental hosted multi-agent model also coordinates server-side subagents.Agents as tools / Handoffs: A powerful mechanism for coordinating and delegating work across multiple agents.openai.github.io/openai-agents-python ↗

Agent permissions

YesPer-tool approval rules pause the run; approvals can be manual interruptions, per-call callbacks, or sticky always-approve decisions, and callable rules fail closed on malformed arguments.Use the human-in-the-loop (HITL) flow to pause agent execution until a person approves or rejects sensitive tool calls. Tools declare when they need…openai.github.io/openai-agents-python/h… ↗

Which models it runs on

Claude

PartlyThe SDK names only OpenAI models; non-OpenAI providers are reached through OpenAI-compatible endpoints or the beta Any-LLM and LiteLLM adapters, and Claude is never named in the docs.It is provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs, as well as 100+ other LLMs.github.com/openai/openai-agents-python ↗

GPT

YesGPT models are the default path: an agent with no model set uses a GPT-5.6 model, and the SDK applies GPT-5-specific ModelSettings.When an Agent does not specify a model, the Agents SDK uses gpt-5.6-luna with reasoning.effort="none" and verbosity="low" by default for…openai.github.io/openai-agents-python/m… ↗

Open models

PartlyAny OpenAI-compatible endpoint or the beta Any-LLM/LiteLLM adapters can be used, which covers self-hosted open-weight serving, but no open-weight model or local runtime is named in the docs.set_default_openai_client is useful in cases where you want to globally use an instance of AsyncOpenAI as the LLM client. This is for cases where the…openai.github.io/openai-agents-python/m… ↗

Model choice

YesYou choose the model per agent, per run, or globally, and can mix models and providers in one workflow.Within a single workflow, you may want to use different models for each agent. For example, you could use a smaller, faster model for triage, while…openai.github.io/openai-agents-python/m… ↗

Where you use it

On the command line

Yesrun_demo_loop gives an interactive terminal chat session with streaming and retained history; it is a demo/testing loop, not a packaged CLI binary.The SDK provides run_demo_loop for quick, interactive testing of an agent's behavior directly in your terminal.openai.github.io/openai-agents-python/r… ↗

In a browser

PartlyNo hosted product UI; the only web surface is OpenAI's Traces dashboard for debugging and monitoring runs.Using the Traces dashboard, you can debug, visualize, and monitor your workflows during development and in production.openai.github.io/openai-agents-python/t… ↗

Not established: In your pipeline

Whose machine it runs on

Self-hosted

YesThe SDK runs in your own Python process, local runtime tools execute in your environment, and sandbox sessions can be Unix-local or Docker on your own machine.Local runtime tools execute outside the model response itself. The model still decides when to call them, but your application or configured…openai.github.io/openai-agents-python/t… ↗

Open source

YesThe repository is MIT licensed, permitting use, modification and redistribution.MIT License Copyright (c) 2025 OpenAI Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated…github.com/openai/openai-agents-python/… ↗

What it costs to run

How it meters

PartlyThe SDK charges nothing; the cost is the model provider's token usage, which the SDK meters per run and per request so you can monitor it.The Agents SDK automatically tracks token usage for every run. You can access it from the run context and use it to monitor costs, enforce limits, or…openai.github.io/openai-agents-python/u… ↗

Free tier

YesThe package is MIT licensed and installed with pip; no vendor account, call or payment is involved, though model calls need an API key from your provider.Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to…github.com/openai/openai-agents-python/… ↗

API access

YesThe product is itself a programmatic surface: a Python package you import and drive from your own application code. There is no separate hosted REST API for it.The OpenAI Agents SDK enables you to build agentic AI apps in a lightweight, easy-to-use package with very few abstractions.openai.github.io/openai-agents-python ↗

MCP server

PartlyThe SDK is an MCP client covering stdio, SSE, Streamable HTTP and hosted MCP servers, with tool filtering and approvals. The docs do not describe exposing your agent as an MCP server for another agent to query.The Agents Python SDK understands multiple MCP transports. This lets you reuse existing MCP servers or build your own to expose filesystem, HTTP, or…openai.github.io/openai-agents-python/m… ↗

Bring your own key

YesYou supply your own API key and can point the client at any base URL or custom OpenAI-compatible endpoint; inference runs on your provider account.By default, the SDK uses the OPENAI_API_KEY environment variable for LLM requests and tracing. The key is resolved when the SDK first creates an…openai.github.io/openai-agents-python/c… ↗

Buying it for a team

Admin controls

PartlyConstraints are code-level, not administrative: approval policies, MCP tool filters and sandbox network policy, all set by the developer.`MCPServerStdio`, `MCPServerSse`, and `MCPServerStreamableHttp` all accept `require_approval`.openai.github.io/openai-agents-python/m… ↗

Audit log

PartlyTracing records every run, tool call, handoff and guardrail by default, but it is developer observability rather than an administrative audit trail, and it is unavailable under ZDR.The Agents SDK includes built-in tracing, collecting a comprehensive record of events during an agent run: LLM generations, tool calls, handoffs,…openai.github.io/openai-agents-python/t… ↗

What happens to your code

Getting out

PartlyRun state serialises to JSON you own and sessions persist to your own SQLite or Redis; the vendor holds nothing except exported traces.`RunState` is designed to be durable. Use `state.to_json()` or `state.to_string()` to store pending work in a database or queue and recreate it lateropenai.github.io/openai-agents-python/h… ↗

Not established: Opt out of training, Data residency, Certifications