OpenAI launches managed Agents API beta
OpenAI launched the Agents API in public beta on September 10, 2026. It provides managed long-running sessions, sandboxes, context handling, tools, and subagent orchestration. Access adds no separate fee, but each task can incur model, tool, sandbox, and third-party costs.
Published 10/9/2026
What happened?
OpenAI made its Codex-based agent harness available to developers through a managed public-beta API. Developers can use an OpenAI-hosted sandbox, their own infrastructure, or an integrated provider. The service supports long sessions, context compaction, tool discovery, programmatic tool calls, and parallel subagents. OpenAI charges no additional Agents API access fee, but underlying tokens and tools are billed. Its usage records are best-effort and are not a final bill.
Why it matters
Freelance developers can prototype long-running, multi-step client workflows without first building the full orchestration layer. Fixed-price delivery still carries margin risk because one outcome can trigger several root-agent and subagent calls, retries, sandbox compute, tools, and third-party services.
What should you do?
Pilot one bounded, non-critical workflow. Set task, tool, concurrency, retry, and spending limits; log cost and failures per completed outcome; require human review for consequential actions; and price from measured end-to-end delivery cost. Keep a fallback because the product remains in public beta.