OpenAI Launches the Agents API, Bringing the Codex Harness to Developers
OpenAI announced the Agents API in public beta on September 10, 2026, positioning it as a way for developers to build and run cloud agents with the Codex harness. The company says the service supplies the managed harness and infrastructure while developers define the task, model, tools, and compute environment. [1]
One API call, several infrastructure choices
The launch example creates a session with the `gpt-6-astra` model, an MCP observability server, up to three concurrent subagents, a vault, an OpenAI-hosted environment, and a task that asks agents to investigate elevated server errors. OpenAI describes this as creating a production-ready agent in a single API call, although production suitability will depend on each application's controls, tools, and deployment requirements. [1]
The API does not require every workload to run in an OpenAI-managed sandbox. Developers can select an OpenAI-hosted environment, their own infrastructure, or an environment supplied through one of the announced ecosystem integrations. OpenAI lists Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel as partners, with deployment, storage, CPU, GPU, memory, cold-start, and cost profiles varying by environment. [1]
| Area | What the announcement says |
|---|---|
| Availability | Public beta |
| Hosted execution | OpenAI-managed sandbox for code, files, and artifacts |
| External execution | Developer infrastructure and listed sandbox partners |
| Coordination | Parallel subagents with configurable concurrency |
| Tool access | MCP, custom functions, built-in web search, and programmatic calls |
The harness handles long-running work
OpenAI says the Agents API automatically compacts earlier context as a session approaches its context limit. That is intended to let an agent continue across multiple context windows without developers implementing their own compaction logic. The company also says its harness is versioned with model launches and maintained as models change, shifting some compatibility work from application teams to the platform. [1]
Tool handling is another central part of the design. Tool search loads relevant definitions when needed, which OpenAI says can reduce token use and cost while preserving the model's cache. Programmatic tool calling allows agents to run calls in parallel, chain operations, and filter or combine results in code before returning selected information to the model. The API supports MCP, custom functions, and built-in tools such as web search. [1]
Subagents become a platform feature
The multi-agent capability lets a primary agent divide a complex task into independent pieces and delegate them to subagents operating in parallel. Each subagent maintains its own context, while the main agent coordinates the work and combines the results. Developers therefore do not need to build all of the orchestration logic themselves, but they still need to decide how tasks are divided, how results are validated, and which actions require approval. [1]
The API reference shows that the beta surface extends beyond agent creation. It documents resources and operations for agents, environments, sessions, artifacts, events, items, turns, subagents, vaults, credentials, and files, including create, retrieve, update, delete, list, and streaming-related operations where applicable. That gives developers an operational surface for inspecting and managing work rather than treating an agent as a single request-response call. [2]
What developers should measure
OpenAI includes customer-reported results as evidence of the API's intended use. Ciridae reported an evaluation-score increase from 0.71 to 0.85 and a fourfold latency reduction; SafetyKit reported a 60% reduction in cost per case; and Hypha reported an 86% reduction in failed agent responses after separating the harness from the sandbox. Those figures are vendor-published customer claims, not independently controlled benchmarks, so teams evaluating the API should reproduce them against their own workloads. [1]
For developers, the practical change is a narrower infrastructure burden but a larger need for operational design. Teams can spend less time implementing context compaction, subagent scheduling, and sandbox plumbing, while investing more attention in tool permissions, artifact handling, failure recovery, observability, and evaluation. The API's public-beta status also makes version tracking and regression testing important before assigning agents long-running or consequential work. [1]