OpenAI Agents SDK Draft
OpenAI · reviewed 22 November 2026 · its card
The OpenAI Agents SDK lets a developer "build agentic AI apps in a lightweight, easy-to-use package with very few abstractions" (S1). OpenAI makes it, as "a production-ready upgrade of our previous experimentation for agents, Swarm" (S1): a Python library, "a lightweight, powerful framework for multi-agent workflows", and a TypeScript one "for multi-agent workflows and voice agents" (S2, S15). It has no surface of its own: "The SDK runs inside your application", whose runner "handles the agent loop and handoffs" (S17). Both libraries are MIT open source (S2, S15), on a 0.Y.Z scheme because the SDK "is still evolving rapidly" (S20). The latest Python release seen is v0.25.0, of 17 November 2026, with the pre-release v1.0.0-alpha.1 of 20 November (S3, S4, S5); the latest TypeScript release is @openai/agents 0.22.0, of 12 November (S16). It is an agent harness. Its maker's own words for it are on the vendor card Harness (OpenAI).
- An instance of
- Agent harness
- Conforms to
- Model Context Protocol (MCP)
Capabilities
| Capability | Value | Detail | Sources |
|---|---|---|---|
| Model | Yes | OpenAI models through the Responses API by default, gpt-5.6-luna unless set; any OpenAI-compatible endpoint or a custom ModelProvider, mixed per agent; Any-LLM and LiteLLM as "best-effort, beta adapter integrations"; local models not mentioned |
S1 S2 S6 |
| Session | Yes | "Built-in session memory to automatically maintain conversation history across multiple agent runs", in SQLite, Redis, SQLAlchemy, MongoDB, Dapr or OpenAI Conversations; a paused run resumes from "the same session ID and the same underlying storage backend" | S7 |
| Subagent | Yes | Agents as tools: "A manager agent keeps control of the conversation and calls specialist agents through Agent.as_tool()"; handoffs instead pass the turn to a specialist |
S8 |
| Agent team | Unknown | The orchestration page describes handoffs, agents as tools, chaining and parallel runs, no agents messaging each other | S8 |
| Skill | Yes | Sandbox agents' Skills capability mounts SKILL.md folders into "a Codex auto-discovery root", .agents by default, eagerly or lazily through load_skill; the pages read do not name the specification |
S10 S12 |
| Instruction file | Unknown | Agents and sandbox agents take instructions in code; no instruction file such as AGENTS.md is named |
S1 S10 |
| Plugin | Unknown | Not found; extensions are optional install extras and third-party adapters | S1 S2 |
| MCP | Yes | Hosted MCP tools through the Responses API, and Streamable HTTP, SSE (deprecated) and stdio servers; "MCP v2 Streamable HTTP supports concurrent requests" from v0.25.0 | S9 S4 |
| Execution environment | Yes | Sandbox agents run in "real isolated workspaces": container through DockerSandboxClient, vendor-sandbox through hosted clients for Blaxel, Cloudflare, Databricks, Daytona, E2B, Modal and Vercel; UnixLocalSandboxClient "adds no OS-level confinement" on Linux, none |
S1 S10 S11 |
| Client | Partial | A library with no window of its own beyond a demo REPL utility, which "can now approve or reject individual tool calls"; realtime agents connect a browser's microphone over WebRTC in the TypeScript SDK | S17 S1 S4 S15 |
| Agent host | Unknown | Not found for other makers' harnesses; v1.0.0-alpha.1 embeds OpenAI's own Codex Core as an opt-in model provider | S5 |
| Channel | Unknown | No channel documented; a realtime session is "a live session object" the application feeds with send_message() and send_audio() while it runs |
S14 |
| Cross-session messaging | Unknown | Not found; coordination is only through handoffs, agents as tools or the application's own code | S8 |
| ACP | Unknown | Not found | None |
| AHP | Unknown | Not found | None |
| A2A | Unknown | Not in the docs read; an open proposal, issue #5660, asks for "Optional A2A support alongside agents.mcp", with no maintainer reply | S18 |
| Bot | Unknown | Not found. Sandbox agent memory "distills lessons from prior runs into files in the sandbox workspace", memories/MEMORY.md; durable long-running runs come from a third party, Temporal's integration, not OpenAI. Scheduling: none documented |
S13 S19 |
State of play
How things stand, newest first, as the card records it.
- Python v1.0.0-alpha.1, a pre-release, "adds Codex as an opt-in model provider for the Python Agents SDK", embedding "unmodified Codex Core
rust-v0.166.0"; wheels only through GitHub Releases. The release page gives no year. S5 - Python v0.25.0 "removes the beta Runloop sandbox provider", lets MCP v2 Streamable HTTP run concurrent requests, and moves
OpenAIResponsesCompactionSessiontogpt-6-luna. S4 - TypeScript v0.22.0-alpha.3, a pre-release, adds typed decisions and app-supplied model backends and widens realtime and voice support; the release page gives no year. S16
- TypeScript
@openai/agents0.22.0, a patch release refreshing dependencies; the release page gives no year. S16 - Google's A2A team opens issue #5660 proposing optional A2A support, offering "to contribute the implementation". S18