OpenAI has introduced the Agents API in public beta, offering developers a managed harness and infrastructure for building long-running AI agents. The API draws on the same orchestration layer that powers Codex and ChatGPT for Work, handling context management, tool orchestration, and subagent coordination so developers can focus on their agent's specific tools and workflows.
The Agents API is available to all developers starting today. There are no additional fees beyond standard token and tool pricing, and OpenAI says it will iterate quickly during the beta period based on developer feedback.
Confirmed
- Managed harness: OpenAI hosts and maintains the agent orchestration layer — context compaction, tool search, programmatic tool calling, and multi-agent delegation — that previously powered Codex internally.
- Environment choice: Developers can run agents in OpenAI-hosted sandboxes, on their own infrastructure (self-hosted), or through one of nine sandbox partners: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
- OpenAI-hosted sandbox: A new fully managed environment that provisions the same sandboxing infrastructure used by Codex and ChatGPT, configurable with files, packages, skills, and plugins.
- Multi-agent support: The API can break complex tasks into independent pieces and delegate them to subagents that work in parallel, each maintaining its own context while the main agent coordinates.
- Open-source foundation: The harness is powered by the open-source Codex codebase (github.com/openai/codex), giving developers visibility into core coordination logic.
- Pricing: No separate API fee; developers pay for tokens and tools used at standard rates (pricing page).
Unknown
- General availability timeline: OpenAI says it will iterate during public beta but has not announced a target GA date.
- SLA and uptime guarantees: No service-level commitments for the hosted sandbox or orchestration layer have been published.
- Self-hosted parity: The extent to which self-hosted deployments receive the same harness improvements on the same cadence as the managed service is not detailed.
- Cost at scale: While per-token pricing is standard, total cost per agent job will depend on context compaction frequency, tool-call volume, and subagent parallelism — workload profiling is needed to estimate real-world spend.
- Partner sandbox feature parity: Whether all nine partner integrations support the full feature set (context compaction, tool search, multi-agent) equally is not specified.
Our take
OpenAI is productizing the orchestration layer that made Codex reliable at scale, turning an internal engineering investment into a developer platform primitive. The partner ecosystem signals OpenAI wants the Agents API to become a neutral control plane across clouds, not a walled garden. What remains unproven is whether the beta's pricing transparency holds at production scale: list price per token is not total cost per agent job when agents run for days, compact context repeatedly, and spin up parallel subagents.
