AI/ES/25/11 Code/LEAD/20 Thu/1100 OpenAI Applied AI Bill Chen Brian Fioca Future-Proof Coding Agents
11:00am - 11:19am Talk: Future-Proof Coding Agents: Building Reliable Systems That Outlast Model Cycles
- Person/Bill Chen LinkedIn - Applied AI, OpenAI
- Person/Brian Fioca Twitter LinkedIn GitHub Website - Applied AI, OpenAI
Talk: Future-Proof Coding Agents: Building Reliable Systems That Outlast Model Cycles 02:11:38 - OpenAI Team (Bill Chen & Brian Fioca) Presentation Starts
- 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.
Agenda
- Coding agents
- Coding agent harnesses
- Codex as both model + harness
- Agents and sub-agents
- Future of Codex
Coding Agents: Core Components
- User Interface (left column; speech-bubble icon)
- CLI
- IDE (e.g., VS Code)
- Cloud / background agent
- Models (center column; robot icon)
- GPT-5.1
- GPT-5.1 Codex Max
- Claude Opus / Sonnet / Haiku
- Composer, SWE 1.5, Gemini, Grok
- Harness (right column; cluster icon)
- Open / closed source
- Prompts
- Agentic loop
- Tools + descriptions: Semantic search, Websearch, Patch/Edit, Browser
Coding Agents: Active Frontier
- Dozens of teams build their own harnesses + prompts
- Tools are diverse/opinionated: search, context-gathering, edits, progress-tracking
- Models keep changing: different strengths, habits, API shapes
- Problem: every model revision forces teams to adapt harnesses + tools
Building + maintaining a harness
- DIY challenges: custom tools out-of-distribution; prompt engineering to fit model; poor prompt portability
- New model challenges: latency; context-window strategy; API response + tool-call format changes
Steerability = intelligence + habit
- Training side-effects: steerable vs baked-in behavior; tool-priming quirks; RL-shaped habits
- Prompts aren’t interchangeable: cloning Claude prompts to GPT-5 hurts latency/reasoning; harness-level steering beats prompt microtuning
03 Codex as a model and a harness
- Codex: many built-in tools
- Parallel tools + shell cmds:
run_terminal_cmd, apply_patch, search - Security + sandboxing: permissions, ports
- Context compaction: prompting, triggers, re-injection, cache tuning
- MCP support: context + security
- Images/screenshots: processing, resizing
- Model is tuned tightly to this harness (and “way harder than you think”).
- Codex bundles capabilities
- Writes + executes code and terminal commands
- Therefore: a computer-use agent for the terminal
- If your task is expressible as file ops + CLI ops, just use Codex
- If your task is expressible as file ops + CLI ops, use Codex
- Even if it doesn’t feel like a coding task
- If it can be done by running tools from the command line, use Codex
- Assign agents to your agents
- Harness = new abstraction layer
- Treat Codex as a pluggable agent inside your agent
- Benefits: stability (swap models safely); prompt decoupling; automatic upgrades
- Result: durable platform that rides the wave, not drowned by it
- Sub-agent patterns
- 01 — Codex SDK
- TS library; programmatic
exec; GitHub Action integration
- 02 — Agents SDK + MCP
- Expose Codex as an MCP tool; give it MCP hooks back into your API
- 03 — Zed ACP
- Instead of IDE-specific harnesses, wrap Codex and let UI/UX pass through
- Direct Codex integration (harness + model)
- Codex SDK supports: CI/CD control; embedding Codex in your app; building agents that talk to Codex directly
- Or directly with the model
- Tweet example (Jediah Katz, Oct 30): CursorAI 2.0 agent-harness improvements boosted GPT-5 Codex quality; runs longer uninterrupted; less overthinking; more accurate edits
- How to do it:
- Check the prompting guide — primary resource for shaping Codex behavior
- Customize only after baselining — reach parity with first-party harness before adding features
- Align with the open-source harness — Codex CLI ships open-sourced prompt + tools for consistency
Future of Codex