OpenAI DevDay: Introducing the Agents API and Building Reliable Agent Infrastructure

Charlie and Steve from OpenAI walk through the reasoning behind the new Agents API, a managed service built on Codex infrastructure that lets developers specify tools, data, and instructions while OpenAI runs the underlying systems. The session covers the engineering challenges of running a locally-designed agent safely in the cloud, keeping long-running tasks observable, and preserving state across deployments, followed by a live rebooking demo built with the API.

■ Evolution of OpenAI's Agent APIs
- From Assistants API to Responses API, and how GPT-6 Astra's long-running tool use shaped the stack
- The layered relationship between Responses API, Codex, and Agents API
■ Core Engineering Problems
- Connecting a cloud agent to arbitrary sandbox or customer-hosted environments
- Giving developers visibility into role-based tool calls and model responses
- Maintaining state and reliability across context compaction, deploys, and server failures
■ Live Demo and Ecosystem
- Airline flight rebooking demo with hosted sandbox, sub-agents, MCP capabilities, and artifact outputs
- Customer examples (Notion, Rakuten, Bin) and managed agents on AWS Bedrock

Targeted at developers building agent products, this session explains what the Agents API manages on your behalf, what design decisions you still own, and how parallel agent workloads change what you can scope for your roadmap.