OpenAI has opened its core agent architecture, which powers the Codex programming assistant, to external developers via the Agents API. The API is now in public preview, allowing developers to access it for free, with pricing based solely on actual token consumption and tool usage. Early customer data shows that the Agents API significantly reduces costs and improves efficiency: SafetyKit reduced per-case processing costs by 60%, Hypha lowered agent response failure rates by 86%, and Cirridae achieved an evaluation score of 0.85 with a fourfold reduction in latency. The API features three key capabilities—automatic context compression, tool discovery and parallel invocation, and multi-agent collaboration—enabling thousands of AI agents to run continuously for hours or even days, meeting the needs of all users, from startup teams to enterprise production deployments.Article author and source: Wall Street Journal
OpenAI is opening its core agent architecture, which powers its flagship programming assistant Codex, to external developers, marking a significant step in the commercialization of its AI agent infrastructure.
On Wednesday, OpenAI announced that the Agents API is now available in public beta for all developers.
This API makes available, as a programmatic interface, the proxy framework and infrastructure previously used exclusively to support Codex and the enterprise version of ChatGPT. Developers can integrate it at no additional cost and are charged only based on actual token consumption and tool usage.
Early customer disclosures show that this product has delivered significant improvements in cost, efficiency, and reliability.
Among them, SafetyKit reduced the cost per case by 60% after migrating its case review workflow to the Agents API; Hypha lowered its agent response failure rate by 86% by decoupling its agent execution framework from the sandbox environment; and technology company Cirridae reported an improvement in its evaluation score from 0.71 to 0.85, along with a fourfold reduction in latency.
Create a production-grade proxy with a single API call.
The core design philosophy of the Agents API is to lower the engineering barrier to building long-running agents.
Developers can create production-ready agents by specifying the task, model, tools, and runtime environment in a single API call, without needing to build underlying infrastructure such as context management or sub-agent orchestration.
This API has three key capabilities:
- First, contextual auto-compression: as the conversation approaches the context window limit, the system automatically compresses earlier content, enabling the agent to run continuously for hours or even days.
- Second, tool search and parallel invocation. Load relevant tool definitions on demand to reduce token consumption, and support parallel execution of multiple tool calls.
- Third, multi-agent collaboration. The main agent can decompose complex tasks and assign them to multiple parallel sub-agents, each of which maintains its own context independently; the main agent then aggregates the results.
Logistics technology company Nash says its platform deploys thousands of long-running AI agents via the Agents API to manage hundreds of millions of delivery tasks worldwide, with agent lifespans spanning hours to days.
Flexible environment selection with ecosystem partners covering diverse needs.
At the computing environment level, the Agents API offers multiple deployment options.
OpenAI provides self-hosted sandboxes that share the same sandbox infrastructure as Codex and ChatGPT, enabling rapid deployment and elastic scaling;
(The Agents API provides AI agents with the same framework and infrastructure as Codex.)
Meanwhile, OpenAI has established first-party integrations with ecosystem partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel, offering developers differentiated options for private cloud deployment, specialized storage mechanisms, and various CPU, GPU, and memory configurations.
This design enables the Agents API to meet the full spectrum of needs, from rapid validation by startup teams to enterprise-grade production deployments.
The AI platform Long Lake, serving the construction and home services sector, says that with this API, its team can now deploy agents across industries in hours rather than the days it previously took.
Open-source frameworks and versioned updates reduce iteration costs.
The Agents API is built on the open-source Codex execution framework; developers can review its public code repository to understand the core logic of model invocation, tool coordination, and context management.
OpenAI maintains and continuously optimizes this framework, allowing developers to automatically benefit from each model upgrade without needing to rebuild their agent architecture.
OpenAI states that the Agents API is currently in public beta and will iterate rapidly based on developer feedback, with plans to move toward general availability.
The company also stated that it supports built-in tools such as the MCP protocol, custom functions, and web search. For developers, there are no additional subscription fees for the API itself—a pricing strategy that lowers the barrier to adoption while directly aligning revenue with actual usage.