Summary
OpenAI’s Charlie from developer experience and Steve from the API team present the Agents API, a managed “harness as a service” built on the same harness that powers Codex. They trace the path from the Assistants API and its hosted File Search and Code Interpreter tools, through the Responses API launched in 2025 and the local Agents SDK, to a hosted service that runs the agent loop for developers.
Under this model, developers define the tools, data, instructions and execution environment, while OpenAI operates the harness and the services around it and keeps improving it as models get better. The speakers explain why that is harder than it sounds: a harness designed for a laptop must run securely in the cloud, developers need visibility into what an agent is doing, and sessions can last hours or days.
The talk covers specific engineering lessons, including exposing turn-level information through the API after an internal team needed it, using compaction to carry history across context windows, routing steering requests to the right running session, and resuming work exactly where it left off after a crash or server deployment. It also references GPT-6 Astra and GPT-6.1 Sol and what these models allow agents to do. It is a useful overview for teams deciding whether to build or buy agent infrastructure.
📺 Source: OpenAI · Published November 11, 2026
🏷️ Format: Keynote Launch
