Introducing the Agents API

# Introducing the Agents API **TL;DR:** OpenAI launches the Agents API, providing developers with a managed agent runtime that handles orchestration, sessions, and context management infrastructure—allowing them to focus on application development. ## Why Do We Need the Agents API? Even with powerful models, deploying a long-running agent to production requires significant effort. Developers must connect tools, track progress, manage context, and secure/maintain the surrounding infrastructure. The Agents API addresses these infrastructure burdens, enabling a more focused development process. ## What is the Agents API? The Agents API brings a managed version of the Codex runtime into your applications. OpenAI handles the underlying orchestration, sessions, and context management, freeing developers from the complexity of building and maintaining their own agent infrastructure. ## Use Case Example: Production Incident Investigation The video illustrates a specific agent application scenario to demonstrate its capabilities. ### Connecting Tools and Context Suppose we want to build an agent that helps investigate production incidents. This agent needs access to observability data and recent code changes while following the team's runbook for handling outages. We can connect all necessary tools via MCP and equip the agent with troubleshooting manuals through a skill module. ### Controlling the Execution Environment You control the agent's execution environment and its accessible tools. This includes connecting sandbox environments via OpenAI, third-party providers, or your own infrastructure. ### Efficiently Handling Large-Scale Data In this scenario, the agent may need to process vast logs that exceed the model's context window capacity. Through programmatic tool calls, the agent can process these logs and filter results via code. This means fewer tokens are used to pass raw data, while more tokens are allocated to the information the agent truly needs—improving efficiency. ## Advanced Capabilities The Agents API also offers a range of advanced features to handle complex tasks. ### Multi-Agent Orchestration For the most complex tasks, work can be delegated to independent agents via multi-agent orchestration. For example, one sub-agent could examine recent code changes, another could analyze telemetry data, and finally, a main agent could synthesize their findings. ### Long Context Management Even within a single agent's session, long-running contexts can work efficiently through a compression mechanism. This mechanism provides the model with a summary of all completed work, allowing the agent to continue seamlessly through its investigation or task. ### Generating Structured Output Once work is done, the agent can generate structured reports. For incident investigations, the final report should include potential root causes, supporting evidence, and recommended next steps—all compiled into a shareable file for the on-call team to review. ## Summary and Vision With the Agents API, developers can drive entire complex workflows without building or maintaining any agent infrastructure of their own. OpenAI states that this underlying infrastructure will continuously improve as models and features evolve. It’s a tool designed to provide developers with a continuously enhanced runtime framework through a single API. Source: https://www.youtube.com/watch?v=2YHa1vhnmK0