OpenAI is opening the Agents API in public beta, turning the orchestration layer behind Codex into a service developers can use to build and run cloud agents. The company says the API is intended for systems that need to work for hours or days, use multiple tools, handle files and code, and preserve enough context to keep a task moving.
The harness becomes the product
The core pitch is a separation of concerns. OpenAI operates the agent loop and maintains the harness; developers supply the model, tools, knowledge, and workflow that make an application distinctive. A session request can define an agent’s model, connect it to MCP or other tools, select an execution environment, and provide the task it should complete.
That is a different proposition from assembling a prompt chain and managing every model call yourself. OpenAI’s example asks an agent to investigate a service’s elevated error rate, delegate deployment, error, and dependency analysis to subagents, and save evidence and recommendations into a workspace. The example is illustrative, but it shows the intended unit of work: a persistent investigation with artifacts rather than a single response.
Where the code runs
Developers can choose an OpenAI-hosted sandbox, their own infrastructure, or a connected environment from a partner. OpenAI lists Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel as first-class integrations, with different trade-offs for networking, storage, compute, cold starts, and deployment control.
The OpenAI-hosted sandbox is designed as the quickest path to a working agent. OpenAI says it provisions the environment, lets developers supply files and packages, and supports skills and plugins so an agent can run code, work with files, and produce artifacts. Teams that cannot place sensitive workloads in a vendor-managed environment can instead keep execution in their own infrastructure while using the hosted orchestration layer.
Long sessions and parallel work
The API packages several pieces of agent infrastructure that are easy to underestimate. OpenAI says automatic context compaction preserves the information an agent needs as a session approaches its context limit. Tool search loads only relevant tool definitions, while programmatic tool calling lets an agent run operations in parallel, chain them, and filter results before returning the useful parts to the model.
Multi-agent support lets a main agent split a complex assignment into focused pieces and coordinate the results. Each subagent keeps its own context, which can reduce the risk that one long task becomes a single, unwieldy conversation. It also introduces a new operational surface: teams will have to monitor concurrency, permissions, tool side effects, and the evidence produced by each branch.
An open-source foundation with a commercial service around it
OpenAI says the Agents API is powered by the open-source Codex harness, giving developers visibility into the core logic that coordinates model calls, tools, and context. The company will operate and improve the managed service as new models arrive, while the public beta gives developers a chance to shape the interface before general availability.
Newsroom analysis
The meaningful shift is that agent reliability is being sold as a reusable platform layer. If the service works as described, developers can spend less time rebuilding compaction, sandboxing, and orchestration, but they also accept a deeper dependency on OpenAI’s runtime choices. The important questions for production teams are therefore not just model quality and token cost; they are where data is executed, how tool permissions are bounded, and how a long-running agent can be audited when its work fans out across subagents.