OpenAI Agents API enters public beta - what developers actually get
Original publication: · Sergey Zhadov
This is a translated archive article. The date and pricing discussion refer to the original publication, not a new announcement today.
OpenAI has released the Agents API in public beta. It gives developers a managed harness for agents that work for long periods, use tools, retain state, and recover from interruptions. This is not a magic autopilot button. Give an agent sloppy permissions and a fuzzy success condition, and it can spend more time and money producing organized rubbish.
What OpenAI actually launched
Agents API provides the working layer that connects a model, tools, context, and an execution environment. A developer describes the agent and its task while the service maintains the session. That can help with repository analysis, document processing, internal operations, and other jobs that do not fit into one short request.
The execution environment is a choice. Teams can use an OpenAI-hosted sandbox, their own infrastructure, or a supported provider. This is not decorative architecture: it affects file access, secret storage, cost, cold starts, and the rules for handling company data.
A long session is not the same as reliability
The harness can compact earlier context, find relevant tools, and split parts of the work between subagents. That removes a lot of plumbing. It also makes a bad initial goal more expensive. One wrong instruction can grow into a whole chain of neatly completed nonsense.
Every task therefore needs a verifiable result, checkpoints, and an explicit stopping condition. Do not ask the agent to improve the project. Name the files, tests, forbidden actions, allowed changes, and report format. The more dangerous the tool, the narrower its permissions should be.
Test it without an expensive circus
Start with ten repeatable tasks and record the human baseline: time, errors, cost, and manual repair. Run the agent in an isolated environment and measure accepted results, not a flashy demo. Count interventions, retries, tool failures, and work sent back for correction.
Break the scenario on purpose. Remove a file, return a tool error, interrupt the network, and provide conflicting data. The agent should stop or recover according to the rules instead of inventing a successful ending. These tests expose the technical crap that a polished presentation usually hides.
Who should try it now
Agents API is relevant to teams that already understand their process and are tired of stitching together sessions, sandboxes, and retries. It may be excessive for one short request. It becomes useful when a workflow involves several steps, files, tools, pauses, and recovery.
The blunt verdict: OpenAI has packaged complicated agent plumbing into a friendlier API, but responsibility has not vanished. Use narrow permissions, a budget cap, logs, tests, and human approval for risky actions. Otherwise the new harness merely helps your existing mess run around the clock.
Translation reviewed:
