OpenAI has opened public beta access to the Agents API, a managed service that provides a Codex-based harness for building production-ready cloud agents with a single API call. The Agents API lets developers define a task, choose a model and tools, and select an execution environment in one request, simplifying the creation of agents that can run long sessions, use external tools, and produce artifacts in secure sandboxes.
The new service supports multiple deployment options so teams can match compute, storage and security needs to their workloads. Developers may run agents in OpenAI-hosted sandboxes, on their own infrastructure, or through integrations with ecosystem partners. OpenAI lists partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel to provide a range of hosting, VPC deployment and storage choices.
OpenAI-hosted sandboxes draw on the same sandboxing infrastructure that powers Codex and ChatGPT. These managed environments are provisioned and maintained by OpenAI and can be configured with files, packages, skills and plugins so agents can execute code, manipulate files and generate artifacts while keeping execution isolated and controlled.
Beyond access to models, the Agents API supplies a versioned harness that coordinates context, tools and orchestration. OpenAI maintains and updates this harness alongside model improvements so developers can inherit new capabilities without rebuilding their underlying infrastructure. The company has emphasized that the managed harness is open source; the core coordination logic is available for inspection on GitHub and documented on OpenAI’s developer site.
A key area of focus is long-running work. The Agents API includes context management designed to sustain sessions that last hours or days. As sessions near model context limits, the system automatically compacts earlier context to preserve the information an agent needs to continue. That automatic compaction removes the need for developers to implement their own compaction logic when building workflows that span multiple context windows.
The harness also aims to improve efficiency around tool usage. The Agents API provides tool search to load only relevant tool definitions when they are needed, reducing token consumption. Programmatic tool calling enables agents to run tool calls in parallel, chain operations and post-process results in code; this model is intended to let agents handle large data volumes while returning only pertinent results to the agent’s context. OpenAI notes the API supports MCP, custom functions and built-in tools such as web search.
To accelerate parallel work, the Agents API includes multi-agent features that allow a primary agent to decompose complex tasks into independent pieces and delegate those pieces to subagents. Each subagent maintains its own context while the primary agent coordinates and aggregates results. According to OpenAI, this architecture can speed research, analysis and coding tasks that benefit from parallelization without forcing developers to build a separate orchestration layer.
OpenAI provides versioned access to harness improvements so agents can benefit from updates to both models and the coordination layer. The company highlighted recent harness enhancements including context compaction, improved tool efficiency and multi-agent support as examples of the kind of capabilities that will be iterated on and delivered through the managed harness.
The Agents API is available in public beta to all developers. During the beta period OpenAI will iterate on the service based on developer feedback. There are no additional fees specifically for using the Agents API; developers pay for the tokens and the tools their agents consume according to OpenAI’s pricing page. For teams that want to inspect or modify how coordination is implemented, OpenAI has published the Codex harness source code and documentation for the Agents API on GitHub and its developer site.
By packaging model access, context orchestration and tool integration into a managed, versioned harness, OpenAI aims to reduce the engineering overhead of building cloud agents while offering deployment flexibility across hosted and partner environments. The public beta provides an opportunity for developers to experiment with long-running agents, tool-enabled workflows and multi-agent designs while OpenAI continues to refine the service based on real-world usage.
Developers interested in trying the Agents API can find the announcement and technical documentation on OpenAI’s site, along with the Codex harness source code on GitHub. As the beta progresses, OpenAI will continue to update the harness and publish improvements so agents built on the platform can take advantage of new features without additional integration work.
Source: Read the original source
Leave a Reply