OpenAI Launches Agents API for Long-Running Cloud Agents
The public-beta Agents API packages session management, tools, sandboxes and multi-agent orchestration into a managed runtime. The bigger decision is where control and responsibility still sit.
OpenAI introduced its Agents API in public beta on September 10, 2026. The service exposes the managed agent harness and infrastructure used behind Codex through an API, targeting workloads that need to keep context, use tools, coordinate subagents and continue for long periods.
What changed
Developers specify a task, model, tools and execution environment. OpenAI operates the harness while teams choose an OpenAI-hosted sandbox, their own infrastructure or an integrated sandbox provider. The release also includes context management for sessions that span multiple context windows, tool discovery, programmatic tool calling and multi-agent coordination.
OpenAI says the public beta has no separate platform fee beyond the models and tools consumed. Public beta still means interfaces and operating assumptions may change before general availability.
Why it matters
Agent projects often begin as a model call wrapped in a loop. Production systems need much more: durable state, recovery, tool authorization, environment isolation, observability and cost controls. A managed harness can remove a meaningful amount of undifferentiated orchestration work.
It does not remove system ownership. The application team still decides what data the agent can read, which actions it may take, how approvals work and what happens when a session fails halfway through a business process.
What teams should evaluate
- Define tool permissions per task instead of giving every agent the same broad access.
- Decide which workloads may run in a hosted sandbox and which require a controlled network or VPC.
- Test cancellation, retry, duplicate actions and recovery from partial completion.
- Capture tool calls, approvals, artifacts, model versions and cost at the session level.
- Plan an exit path for critical state so the business process is not inseparable from one harness.
BronzeDev view
The release shifts the build-versus-buy boundary upward. Teams can spend less time maintaining an agent loop, but integration engineering remains the difference between a useful agent and an unsafe automation. Start with one bounded workflow and measurable completion criteria before expanding tool access.
