OpenAI's Agents API turns the execution layer around a long-running agent into a managed service. The operator still owns the harder production choices: environment, authority, recovery policy, evidence, and acceptable cost.
What Entered Public Beta
OpenAI announced the Agents API on September 10, 2026. A developer creates a session with a task, model, tools, environment, and optional multi-agent configuration. OpenAI hosts and maintains the Codex-derived runtime that coordinates model calls, context, tools, and subagents.
The launch says the public beta is available to all developers. Usage carries the token and tool charges for the work performed, with no separate Agents API fee.
The Runtime Boundary Is Selectable
The product documentation separates the agent definition, session, and execution environment. An agent can run in an OpenAI-hosted sandbox, on infrastructure the builder operates, or with an integrated sandbox provider. The listed provider options include managed environments and VPC deployments with different file, secret, compute, latency, and cost profiles.
That makes the environment a declared dependency instead of an invisible property of the assistant. A production design should record the environment type, image or dependency version, egress policy, secret boundary, and artifact destination with every run.
Context And Delegation Move Into The Service
OpenAI says the runtime can compact earlier context as a long session approaches its limit, load tool definitions on demand, execute related tool calls in code, and delegate independent work to subagents with separate contexts.
Those features reduce infrastructure a builder must write. They also move more runtime policy behind a versioned provider boundary. Store the runtime version, compaction events, delegated run graph, tool inputs, approvals, outputs, and costs outside the prose transcript when those records affect audits or recovery.
What The Announcement Does Not Prove
- Public beta does not establish a reliability target for a specific workload.
- Selected customer statements are not independent performance evaluations.
- A hosted sandbox does not define which business actions an agent may take.
- Automatic compaction still needs checks for lost constraints and evidence.
- Multi-agent parallelism can increase cost, side effects, and review load.
Operator move: Run one recoverable workflow through two environment options. Compare completion rate, intervention rate, duplicated side effects, evidence retained after compaction, cost, and time to resume after a forced interruption.