Future-Proof Coding Agents: Building Reliable Systems That Outlast Model Cycles

Time: 11:00 AM

Speaker Bios: Bill Chen is Product Manager at OpenAI. Brian Fioca works in Engineering at OpenAI.

Speaker Profiles: Bill Chen | Brian Fioca

Company: OpenAI creates GPT models and coding assistants, including advanced reasoning capabilities.

Focus: How to build coding agents that remain reliable as models evolve. Critical for understanding production-grade agent architecture.

Reference: YouTube Link

Slides

Slide: 10-22

Key Point: Trust in AI grows through predictability and understanding, allowing users to delegate larger tasks with less oversight. The concept of “fingerspitzengefühl” (intuitive feel) becomes important in human-AI collaboration for navigating complex situations.

Literal Content:

• Title: “Gene’s Hopes And Favorite Findings”
• DORA logo in top right
• Pink background
• Three bullet points:
  • “Observation: as I’ve worked with AI more, I trust the AI to do larger tasks”
  • “One strange definition of trust: to what degree do can I predict how another party will act and react — the more I trust the other party, I can make bigger requests, with fewer words, with less feedback”
  • “The notion of fingerspitzengefühl: person’s instinctive ability to handle complex, uncertain situations with intuition, tact, and sensitivity (thank you Idan Gazit, GitHub)“

Slide: 10-25

Key Point: This is a resource slide offering extensive materials on DevOps, AI development, and organizational transformation, inviting attendees to email for access to Gene Kim’s research and publications.

Literal Content:

• Title: “Want More Learn More?”
• Left side lists resources including:
  • Excerpts of Vibe Coding and Wiring the Winning Organization
  • Updates on benchmarking GenAI and developers
  • Best talks from ETLS Community
  • PDF and audio excerpts from The Unicorn Project
  • Eight excerpts from Beyond The Phoenix Project audio series w/John Willis
  • 140 page excerpts from The DevOps Handbook and The Phoenix Project
  • Videos and slides from DevOps Enterprise 2014-2019
  • One hour excerpt of The Phoenix Project audiobook
• Contact information: “[email protected]” with subject “vibe”
• @RealGeneKim handle

Slide: 11-15

Key Point: OpenAI is presenting three architectural patterns for integrating Codex as a sub-agent: direct SDK integration, MCP-based tool exposure in agent frameworks, and IDE wrapper integration through Zed ACP. This demonstrates multiple integration strategies for different use cases.

Literal Content:

• Title: “Sub-agent patterns”
• Pink background
• Three numbered sections:
  • “01 Codex SDK” - “Codex can be called through a Typescript library, programmatically via exec, or as a GitHub Action.”
  • “02 Agents SDK + MCP” - “You can expose Codex via MCP as a tool in Agents SDK, as well as give it MCP hooks to call back to your API”
  • “03 Zed ACP” - “Instead of building a harness in your IDE, wrap Codex and pass through UI/UX”
• Footer: “OpenAI | Confidential and proprietary.”

Notes

• Ground is shifting so fast
• Agents
  • Agents
  • Harnesses
  • Agents and subagents
• Talking about Codex specifically
• 3 parts
  • Interface
  • Models
  • Harness (focus on today)
    • Prompts
    • Agent loop
    • Tools + tools descriptions
      • Semantic search
      • Web search
      • Patch / edit
      • Browser
• “Hard to track the models and we aren’t making the problem easier for anybody” people need to adapt to the new models
• Harness
  • The surface area that the model uses to talk to the user and code and interact with tools
  • For some the harness might be the special sauce of the product
  • Challenges
    • Custom tools be out of distribution (doesn’t know how to use it)
    • Prompt engineering needs to fit in with how to use the tool
    • Poor portability of prompts across models
    • Latency -> context management -> API
      • Codex Map should do that for you
      • https://openai.com/index/gpt-5-1-codex-max/
    • Steerability = intelligence + habit
      • Training has side effects
      • E.g. apply patch quicks
      • Prompts aren’t interchangeable
      • Harness driving steering > prompt microtuning
• “I like the solution that you came up with but it took too long to come up with, what can I do to make it better”
• Harness + model combined
  • Many things under the hood
  • Parallel tools
  • Security and sandboxing
  • Context compaction
  • MCP support
  • Images and screenshots
• Examples
  • Use Codex to organize photos into a folder
  • Analyze a huge amount of CSV files in the terminal
• Use Codex the agent inside of your own agent
  • Durable platform that rides the wave instead of drowning in it
• Can be called through the SDK, agents SDK + MCP + Zed ACP
  • Can build out software that it needs that it doesn’t have
  • Zed wraps Codex into a layer
• You can customize the coding agent
  • Align the tools to be in distribution of how it was trained
• Dozen of trillions of tokens per week
• Build where the models are going
• New models will raise the trust ceiling