From FizzBuzz to 10,000 Agents: The Evolution of AI Agents at OpenAI

OpenAI's Applied AI engineering team and DX team walk through the evolution of AI agents — from early ChatGPT code generation to today's Codex — and share practical design patterns and upcoming directions. The session covers real-world workflows, subagent orchestration, and a large-scale research effort.

■ Agentic Loop & Harness Design
- From ChatGPT's FizzBuzz demo to o3's tool use and the agentic loop
- Codex architecture: model + harness, Responses API, and sandbox execution
- Context compaction and long-horizon persistence across GPT-5, 5.5, and 5.6

■ Working with Codex Today
- Starting with a single agent vs. adding subagents
- Effort scaling with tiered agent roles (Sol, Astra, Luna)
- Pair agents, thread management, and work logs for staying on track
- Multi-agent coordination with GPT-6

■ Scale & What's Next
- Navier-Stokes problem: 10,000 agents, 2.7 million messages, 130 billion output tokens
- Agent team architecture: unified goals, cross-team communication, and avoiding duplicated effort
- Next-generation agent capabilities and how to apply these approaches to your own projects

エンジニアやAIエージェント開発に関心のある方向けに、Codexの実践的なワークフローやマルチエージェントの設計パターンが具体的な事例とともに紹介されています。視聴後は、単一エージェントから大規模チームまで、自プロジェクトに応じたエージェント活用の進め方を具体的にイメージできるようになります。