The Pulse
OpenAI Opens Managed Agents API to All Developers
OpenAI has placed its Agents API into public beta, giving developers access to a managed Codex harness for long-running, tool-using agents. The service supports durable sessions, hosted or third-party sandboxes, MCP connections, context com
AI.info Team ·
“With the Agents API, our evaluation score went from 0.71 to 0.85.”
Jack Weissenberger, chief technology officer, Ciridae
OpenAI has opened its Agents API to all developers in public beta, offering a managed version of the Codex harness that handles the infrastructure behind long-running software agents. The company says developers can use the service to coordinate model calls, tools, files, context and subagents without building and maintaining those systems themselves.
OpenAI announced the release on September 10, 2026, alongside documentation, a quickstart and a new hosted sandbox. The launch gives developers a choice between an OpenAI-managed execution environment, infrastructure they operate themselves, and integrations from a group of outside providers.
The product targets a specific problem in agent development: keeping an AI system working across multiple steps, tools and context windows while preserving enough state to continue after delays, failures or interruptions. OpenAI describes the service as the same harness and infrastructure that powers Codex, adapted for use through an API.
OpenAI Takes Over the Agent Loop
Agents API separates the agent harness from the environment where the agent runs. OpenAI operates the harness, including session orchestration, context management and recovery. Developers choose the model, tools, instructions, data and execution environment.
That division lets a developer create an agent session with a single API call. OpenAI’s example specifies a model, an MCP server, a vault, an OpenAI-hosted environment and an input task that asks the agent to investigate a service’s elevated 5xx error rate. The task directs the system to delegate deployment, error and dependency analysis to subagents, then save evidence and recommendations in a workspace directory.
The API supports durable sessions that continue across turns and can stream progress while the agent works. Developers can connect custom functions, MCP servers and built-in tools such as web search. The service also supports reusable agent configurations that can be retrieved by project through the beta API.
OpenAI’s release notes describe the product in direct operational terms: the company handles session orchestration, context compaction and recovery, while developers connect their own tools and MCP servers. The arrangement reduces the amount of state-management code that teams need to write before an agent can handle a task that lasts longer than a single model response.
Hosted Sandboxes Sit Beside Developer Infrastructure
OpenAI’s hosted sandbox is designed for agents that need to run code, work with files and produce artifacts. The company says it provisions and manages the environment, which can be configured with files, packages, skills and plugins required by a particular workflow.
Developers are not required to run their agents in that environment. OpenAI lists Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel as ecosystem providers with first-class integrations. The company says those options cover different deployment arrangements, including fully managed environments and deployments inside a customer’s virtual private cloud.
The provider choice also affects storage, secrets, compute, memory, startup time and cost. OpenAI presents the environment as an application decision rather than a fixed part of the API. That matters for teams whose agents need access to private data, specialized compute or existing internal systems.
OpenAI’s documentation makes the division explicit: the company maintains the harness, while the developer controls where the agent runs code and handles files. A customer can therefore use OpenAI’s orchestration layer without placing every execution step inside an OpenAI-operated sandbox.
Long Sessions, Tool Search and Subagents
The service’s main technical features focus on the parts of agent systems that tend to become difficult as tasks grow. OpenAI says its context-management layer automatically compacts earlier material as a session approaches its context limit. The agent can then continue across multiple context windows without requiring the developer to implement a separate compaction system.
Tool search loads relevant tool definitions when they are needed instead of placing every available tool in the model’s context at the beginning of a task. OpenAI says the approach can reduce token use and cost while preserving the model’s cache. Programmatic tool calling allows an agent to run calls in parallel, chain operations and filter or combine results in code before returning selected information to the model.
Multi-agent support lets a main agent divide a complex assignment among independent subagents. Each subagent maintains its own context, while the main agent coordinates the work and combines the results. OpenAI’s code sample enables multi-agent execution with a maximum of three concurrent subagents, although the company’s customer examples describe larger deployments.
OpenAI does not present the customer figures as independent evaluations. The company quotes Ciridae’s chief technology officer, Jack Weissenberger, saying the firm’s evaluation score increased from 0.71 to 0.85 and that the new API produced a fourfold reduction in latency for its subagent flows. Other customer statements cite a 60% reduction in cost per case at SafetyKit, an 86% reduction in failed agent responses at Hypha and the ability to fan work across hundreds of agents at Dwelly.
Those numbers describe individual customer deployments and conditions, not a common benchmark. Their importance lies in the problems they identify: orchestration overhead, cost per task, latency and failure recovery. OpenAI is positioning the product around those operating details rather than around a new model release.
Codex Becomes an API Foundation
OpenAI says Agents API is powered by the open-source Codex harness. Developers can inspect the public codebase, while OpenAI continues to operate and improve the managed version used by the API.
The company also says the API provides versioned access to harness capabilities alongside model launches. That arrangement is intended to reduce the need for developers to rebuild their orchestration layer whenever a new model changes how it handles tools, context or multi-step work.
Codex began as a coding-focused product, but OpenAI’s API examples point to broader uses. An agent can investigate an operational incident, search connected knowledge, write files, run code, call external services and delegate parts of the task to other agents. The same execution pattern can support software operations, document review, logistics, research or internal business workflows, provided the developer supplies the relevant tools and permissions.
One customer statement comes from Aziz Alghunaim, co-founder and chief technology officer of Nash.ai. OpenAI quotes him saying the company uses the API’s durable-session and orchestration layer for agents that manage hundreds of millions of deliveries across global logistics networks. Nash supplies the tools and execution environment, while the API handles context, recovery and multi-step coordination.
The split resembles a managed runtime more than a conventional text-generation endpoint. Developers still define what the agent is allowed to do, which systems it can access and where it can execute actions. OpenAI takes responsibility for the machinery that keeps the workflow moving.
Public Beta Has No Separate API Fee
Agents API is available in public beta to all developers. OpenAI says it does not charge an additional fee for using the service. Customers pay for the tokens and tools their agents consume under the company’s existing pricing rules.
That pricing structure lowers the initial barrier to testing the product, but it also makes usage design important. A long-running agent that calls several tools, maintains multiple sessions and spawns subagents can consume substantially more resources than a single response request. Tool search and programmatic calling are intended to control some of that overhead, while the choice of sandbox affects compute and storage costs.
OpenAI has not declared a general-availability date. The company says it will iterate during the beta as developers provide feedback about production use, friction and missing capabilities. The public release therefore establishes the API surface and operating model, but not a final contract for every feature.
The immediate change for developers is practical. Teams no longer need to assemble every part of an agent runtime before testing a long-running workflow: session state, context compaction, tool discovery, recovery and subagent coordination are available through one OpenAI-managed layer. The remaining work sits where it always does in production systems—choosing the right tools, setting permissions, selecting an execution environment and deciding what the agent may do without human approval.