01
Overview
OpenAI launched the Agents API in public beta on September 10, 2026, giving developers managed access to the same evolving agent harness that powers Codex. The important change is not simply another endpoint: OpenAI is packaging orchestration, long-running session management, context compaction, tool discovery, programmatic tool calling and multi-agent coordination behind an API while letting developers choose where agent code runs.
02
What is the OpenAI Agents API?
The Agents API is a managed cloud-agent API. A developer defines the task, model, tools and environment, and OpenAI runs the agent loop. OpenAI says the harness is based on the open-source Codex harness, while the hosted service maintains the orchestration layer for developers.
The API is in public beta and available to all developers. OpenAI says there is no separate fee just for using the Agents API; usage is billed through the models and tools the agent consumes. Hosted sandbox/container usage can add infrastructure charges under OpenAI's container pricing, so 'no Agents API fee' should not be read as 'agents are free to run.'
03
Why this launch matters
Agent applications usually need more than a model call. They need a loop that can keep state, select tools, recover from partial work, manage long context and coordinate work across multiple steps. OpenAI is moving those responsibilities into a maintained harness rather than requiring each developer to rebuild them.
That makes the Agents API especially relevant for coding, research, operations, data analysis and business workflows that can run for minutes, hours or potentially days.
04
Long-running sessions and automatic context compaction
OpenAI says the API can support workflows spanning multiple context windows. As a session approaches a model's context limit, the harness can automatically compact earlier context while preserving information needed to continue.
This is important for long tasks because it reduces the amount of custom memory and summarization logic developers need to implement themselves. It does not remove the need to design durable application state, audit logs or domain-specific memory when those are required.
05
Tools: MCP, functions and built-in capabilities
The Agents API supports MCP servers, custom functions and built-in tools such as web search. OpenAI also describes tool search, which loads relevant tool definitions when needed instead of forcing the full tool catalog into every model turn.
Programmatic tool calling can run operations in parallel, chain calls and filter or combine tool output in code before returning only relevant information to model context. For tool-heavy agents, that can reduce unnecessary context usage.
06
Multi-agent and subagent support
The API supports a multi-agent mode in which a main agent can delegate independent pieces of a task to subagents. Each subagent has its own context and the main agent coordinates the results.
This is useful when a task can be decomposed cleanly: for example, one subagent investigates deployment state, another analyzes logs and another checks dependencies. It is not automatically better for every workflow; parallelism adds cost and coordination overhead when subtasks are tightly coupled.
07
Choose where the agent works
OpenAI separates the harness from the execution environment. Developers can use an OpenAI-hosted sandbox, their own infrastructure, or supported sandbox providers.
OpenAI lists ecosystem integrations with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. These options can differ in CPU/GPU resources, storage, networking, cold-start behavior, VPC deployment and cost.
08
OpenAI-hosted sandboxes
OpenAI's hosted sandbox gives agents an environment to run code, work with files and generate artifacts. Developers can provide files, packages, skills and plugins. OpenAI provisions and manages the environment using infrastructure related to what powers Codex and ChatGPT.
Hosted sandbox use should be budgeted separately from model tokens. Exact container rates, resource classes and limits should be checked against the current OpenAI pricing and developer documentation before production deployment.
09
Pricing
OpenAI states that the Agents API itself adds no separate API fee. Developers pay for the model tokens and tools their agents use. Any hosted sandbox/container costs, external MCP services, third-party APIs, storage or network services can add to the total cost.
A practical total-cost model is therefore: model usage + OpenAI tool usage + sandbox/container usage + external APIs/services + your own infrastructure.
10
What the Agents API does not remove
The service can reduce orchestration work, but developers still need to define permissions, credentials, environment boundaries, task design, evaluation, observability, failure handling and user experience.
The public beta designation also matters: request formats, supported models, rate limits, integrations and behavior can change. Production teams should pin supported versions where possible and reverify documentation before deployment.
11
Agents API vs Agents SDK
The Agents SDK is a developer library for building agent logic. The new Agents API is a managed service that runs the agent harness for you. They overlap conceptually, but the key distinction is operational responsibility: the API moves more of the durable orchestration and harness operation to OpenAI.
Developers who need deep control or a fully self-operated stack may still prefer SDK-based or custom orchestration. Teams that want to ship long-running cloud agents without maintaining the full loop themselves are the clearest initial fit for the Agents API.
12
Likely use cases
- Long-running coding and repository tasks
- Research agents that split work across sources
- Operations and incident investigation
- Data analysis pipelines with files and code execution
- Business workflows combining MCP tools and internal systems
- Multi-step automation requiring durable sessions
- Agent applications that need generated artifacts
13
Availability
OpenAI says the Agents API is available in public beta to all developers as of September 10, 2026. General availability timing has not been announced.
14
FAQ
Is the OpenAI Agents API free?
There is no separate Agents API fee according to OpenAI, but you still pay for model tokens and tools, and hosted sandbox/container usage can add cost.
Does it support MCP?
Yes. OpenAI documents MCP support alongside custom functions and built-in tools such as web search.
Can it run multiple agents?
Yes. Multi-agent support allows a main agent to delegate work to parallel subagents.
Can I use my own infrastructure?
Yes. OpenAI says developers can use OpenAI-hosted sandboxes, their own infrastructure, or supported environment providers.
Is it production-ready?
It is publicly accessible, but OpenAI labels it public beta. Treat interfaces and limits as subject to change until GA.
15
Bottom line
The Agents API turns the Codex-style harness into a managed developer primitive. Its strongest value proposition is not a new model but reduced orchestration burden for long-running, tool-using and multi-agent workflows. The biggest questions to monitor next are real-world costs, rate limits, environment constraints, reliability at scale and how quickly the public beta stabilizes.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.