Summary
In this DevDay 2026 session, OpenAI Applied AI engineer Raj Bhatek and developer experience engineer Eric trace how agent design has evolved and what comes after the simple agentic loop. Raj starts with the first ChatGPT FizzBuzz example, moves through o3’s combination of reasoning and tool use, and explains how Codex was built around a model and harness trained to work together.
The talk describes how the harness manages instructions, tools, context and the execution environment, with the Responses API at the center of the loop. It then follows the progression from GPT-5 handing over whole tasks, to GPT-5.6 with persistence trained directly into the model, to the GPT-6 series, which can compact context, push through failures and coordinate its own sub-agents without an external loop.
Eric offers practical guidance for developers: start with a single agent and a clear way to check its work, and add more agents only when the problem genuinely decomposes. Using Ultra, Codex can spawn copies of itself for large projects such as code-base refactors, but every added agent can sharply increase token use without making the work faster or better. The session is a useful look at the design decisions behind OpenAI’s agent stack.
📺 Source: OpenAI · Published November 11, 2026
🏷️ Format: Keynote Launch
