OpenAI Launches the Agents API in Public Beta
OpenAI has officially launched the Agents API in public beta, providing developers with a managed service to run long-lasting AI agents.
Stock photo for illustration only, not from the actual event
- OpenAI launches the Agents API in public beta, live for all developers.
- Built on the open-source Codex harness for managing context and subagents.
- Supports 3 sandbox options along with a non-sandbox execution mode.
- Data remains strictly US-only with Zero Data Retention currently unsupported.
OpenAI has officially rolled out its new Agents API into a public beta phase, making the managed service live and accessible for all developers to deploy in production environments. According to the announcement, all user data will remain strictly within the US, and Zero Data Retention is currently unsupported.
The Agents API is built as a managed service on top of the open-source Codex harness. The OpenAI team noted that scaling Codex and ChatGPT for Work highlighted the core requirements of long-running agents, which include a robust harness to manage context, utilize tools efficiently, coordinate subagents, and maintain reliable execution over multiple days.
The introduction of the Agents API marks a crucial shift for OpenAI from merely offering raw AI models to delivering comprehensive application infrastructure. By abstracting the complex orchestration required for autonomous agents, OpenAI is directly addressing the developer bottleneck in building multi-step, persistent AI applications.
Official documentation organizes the API around four core concepts. A typical session follows a structured workflow designed to handle complex programmatic tasks efficiently:
The official documentation organizes the API around four distinct concepts. A session execution flow operates across four key steps:
- Create a session and assign a specific task to the environment
- Monitor ongoing progress in real-time via streaming or webhooks
- Continue operations with a new task or steer the current turn dynamically
Stock photo for illustration only, not from the actual event
When it comes to architectural decisions, choosing the right environment is paramount. The Agents API natively supports three distinct sandbox options, while also offering the flexibility to execute workflows without utilizing any sandbox environment.
Furthermore, OpenAI maintains the underlying harness alongside its foundational models, ensuring versioned access with each model launch to maintain long-term stability and developer compatibility.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment