What Comes After Loops?: Deep Dive Into Agentic Architecture | DevDay 2026
Summary
OpenAI's evolution from ChatGPT copy-paste to Codex agents reveals a critical shift: agents aren't just better models—they're model-harness co-design systems where training and infrastructure inform each other. The real breakthrough is giving models computer access and reasoning loops that let them self-correct, automating entire workflows vs. single outputs.
Key Takeaways
- The agentic loop combines reasoning + tool use, enabling AI to act, observe results, and adjust course automatically—shifting from static outputs to dynamic problem-solving that mirrors human iteration.
- Codex's harness design manages instructions, tools, context, and execution environment as a unified system. The model and harness co-evolve: harness design informs training, and model capabilities inform harness architecture.
- Native context compaction built into both harness and model prevents agents from losing previous work during long tasks—critical for maintaining state across multi-step workflows that exceed single context windows.
- Minimalist, generalizable tooling (shell, file system, computer use) enables agents to generalize across tasks like humans do, rather than requiring bespoke tools for each use case.
- Agent architecture has evolved rapidly in 1 year—what works today may be outdated soon as models improve, forcing builders to prioritize flexible harness design over rigid agent frameworks.
Related topics
Transcript Excerpt
Hey everybody. How's it going? Great. My name is Raj Bhatek. I'm part of our Applied AI engineering team here at OpenAI, and excited to be with you all today. When we were naming the session, What Comes After Loops, we joked that by the time we finished the content, everything would be outdated. So what we have today, it's for today. But that's honestly what building agents has been like for the past year or so. As models get better, things change. How we build them really does, too. So today, Eric and I are going to rewind a bit, tell you about how we got here, some of the design decisions we made here at OpenAI, what you can do with agents today, and then share some stuff that we've been cooking. So before we get started, I want to rewind a bit. I want to take you guys back to when we la…
More from OpenAI
- How to request and enable Daybreak Blue
- Build What’s Next on the ChatGPT Platform with Plugins | DevDay 2026
- How Mozilla uses GPT-6 Cyber to find and fix Firefox vulnerabilities
- Building with GPT-Live and Image Generation | DevDay 2026
- Harness-as-a-Service: Building with the Agents API | DevDay 2026