Agents API
A managed agent development service introduced by OpenAI, powered by the Codex harness, that supports the construction and deployment of cloud-based agents. Core capabilities include task orchestration, long-running sessions, and native tool calling.
Core Facts
在MCP出现之前，AI调用外部工具主要依赖OpenAI的Function Calling机制或各框架自定义的插件系统
The technical foundation of the Agents API is powered by the Codex harness.
The Agents API provides native support for long-running sessions, enabling agents to handle tasks that take minutes, hours, or even longer to complete.
Timeline (last 90 days)
Computer Use is built into the Agents API, so developers can directly call the same implementation as Codex and ChatGPT, avoiding the high barrier of building their own harness.
Google built the Agents API on top of the sandbox, with the core idea of defining an agent's configuration (underlying model, instructions, environment) as a reusable, shareable entity, where the environment can be raw (with source files) or an existing agent or environment.
OpenAI launched the Decisions API at DevDay, powered under the hood by a dedicated gpt-6-luna model.
By combining the Agents API with cloud-based computer operations, developers only need a single platform project and a single API key to build agent applications with computer operation capabilities, running in an OpenAI-hosted browser.
The Agents API runs a managed codex harness, providing capabilities such as persistent sessions, orchestration, context compression, and fault recovery
The core criterion for selecting an approach is whether the data control requirements are compatible with the managed solution; if data residency and retention policies meet the requirements, the Agents API is one of the fastest paths to complex general-purpose agents
The Agents API is currently not eligible for ZDR (Zero Data Retention)
The Agents API currently supports US residency only
The Agents API has entered preview, supports computer use, and is the same technology powering all cloud-based agents, including Dots.
Computer use capabilities allow agents to open a browser and operate a graphical interface to complete real tasks such as end-to-end testing.
24 more timeline events
All Facts (20)
在MCP出现之前，AI调用外部工具主要依赖OpenAI的Function Calling机制或各框架自定义的插件系统
80%VerifiedThe technical foundation of the Agents API is powered by the Codex harness.
75%VerifiedThe Agents API provides native support for long-running sessions, enabling agents to handle tasks that take minutes, hours, or even longer to complete.
75%VerifiedThe project adopts a multi-agent collaborative architecture, decomposing job-seeking tasks into independent agents with different responsibilities.
75%VerifiedOpenAI's Agents API has entered preview, providing underlying support for all cloud agents including Dots, and also supports computer operations.
65%UnverifiedComputer Use is built into the Agents API, so developers can directly call the same implementation as Codex and ChatGPT, avoiding the high barrier of building their own harness.
50%UnverifiedThe core criterion for selecting an approach is whether the data control requirements are compatible with the managed solution; if data residency and retention policies meet the requirements, the Agents API is one of the fastest paths to complex general-purpose agents
50%UnverifiedThe Agents API runs a managed codex harness, providing capabilities such as persistent sessions, orchestration, context compression, and fault recovery
50%UnverifiedComputer use capabilities allow agents to open a browser and operate a graphical interface to complete real tasks such as end-to-end testing.
50%UnverifiedTeams with their own in-house agent frameworks need to evaluate the compatibility and vendor lock-in risks of migrating to the official API.
50%UnverifiedArtifacts generated by a task do not automatically appear on the local machine; the application must explicitly call the Artifact API to first obtain an Artifact ID and then call the Content interface to download it.
50%UnverifiedThe way a Skill enters an API request is by packaging the entire Skill directory into a zip file, with the skills array using the base64 type, media_type set to application/zip, and data set to the zip-encoded content.
50%UnverifiedThe uploaded Skill is mounted to the path .managed_agents/.skills/CSV-report in the workspace.
50%UnverifiedThe entry point for the Agents API is /v1/agents/sessions, and the request header must include an API Key, the OpenAI-Beta: agents=v1 flag, and the application/json content type.
50%UnverifiedInline files contain three key fields: type is inline, path determines the file's location in the sandbox, and data stores the Base64-encoded file content.
50%UnverifiedThe Agents API has three levels: a Session is a persistent task container, a Turn is a single advancement or iteration within it, and Items are the messages, commands, tool results, and final outputs generated during an advancement.
50%UnverifiedWhen stream: true is set, the Agents API server continuously returns structured events via SSE, from Agent Session Created through Turn Completed and Session returning to Idle.
50%UnverifiedThe Agents API is not a model, but a runtime framework built around models.
50%UnverifiedIn the Agents API, a team's runbook is provided to the Agent in the form of a skill.
50%UnverifiedIn the Agents API, control of the execution environment is held by the developer, and the sandbox can use infrastructure provided by OpenAI, third parties, or the developer themselves.
50%