← back
Future-Proof Coding Agents – Bill Chen & Brian Fioca, OpenAI
Takeaway
Don't port prompts across models — match instructions to the model's trained habits, or use a co-developed model+harness like Codex.
Summary
- Bill Chen & Brian Fioca (OpenAI applied AI startups) break coding agents into three parts — interface, model, harness — and argue harnesses need to be rebuilt or re-tuned each model release.
- Key insight: 'intelligence + habit' — every trained model has habits (e.g., GPT-5 plans, gathers context, tests before edits); over-prompting it like an older model causes it to over-examine files and slow down.
- Recommends asking the model directly what's wasting time in its instructions — they literally asked and the model said 'you're telling me to look at everything I don't need to'.
- Codex (just shipped GPT-5.1-Codex-Max) bundles model + harness: handles parallel tool calls, sandboxing, port mgmt, automatic compaction, MCP plumbing, image resolution — all the hard parts.
- Frames Codex as a 'computer use agent for the terminal' — anything expressible as files + commands (CSV analysis, photo organization) is in-scope, not just coding.
coding-agentscodexopenai
Original description
Coding agents are becoming one of the most active areas in applied AI, yet many teams keep rebuilding fragile infrastructure every time models or providers change. We believe there is a better way. By anchoring on a stable abstraction layer like Codex, we can stop worrying about harness rewrites and focus on the parts of the stack that create lasting value. We treat models as interchangeable sub-agents, plug into shared primitives, and let upstream improvements flow through without breaking products. This lets teams move faster, stay resilient as the ecosystem evolves, and focus their energy on domain-specific workflows and user experience. Speakers: - https://twitter.com/bfioca - https://twitter.com/realchillben AIE is coming to London and SF! see https://ai.engineer for dates and sign up to be notified! \**Timestamps:** 00:00 Introduction & The State of Coding Agents 02:06 Anatomy of a Coding Agent & The "Harness" Definition 03:44 Technical Challenges in Building Harnesses 06:03 Intelligence vs. Habit: Lessons in Prompt Engineering 08:24 Deep Dive: Codeex as a Harness/Agent 10:33 Computer Use & Terminal Capabilities 11:25 Patterns for Building with the Codeex SDK 14:32 Case Studies: GitHub & Cursor Integration 15:34 Future of Coding Agents & Conclusion