Codex: Why Cheaper Rewrites Still Need Architecture (Category #AI4SDLC)
Agents make systems cheaper to rework, but good boundaries and architecture become even more important. Teams used to grow gradually, developing shared agreements and documentation along the way. Now, in Tibo Sottiaux's illustrative comparison, a hundred agents could join a project over a weekend. And all of them will start changing things. By then, it would help to understand what we are building and how those changes are supposed to coexist.
This is an interesting thread in Building Codex with Tibo Sottiaux, an interview published on September 9 by The Pragmatic Engineer. Tibo is one of the engineers who created Codex and now leads Core Products & Platform at OpenAI. His interviewer is Gergely Orosz, whose talk about slowing down to speed up I covered earlier.
Tibo has a consistent interest in tools that help other people work faster. At DeepMind, he worked on research infrastructure. He describes how an interface for experimenting with language models grew into an internal chat application that colleagues were enthusiastically sharing before ChatGPT launched. It never became a public product; Tibo attributes that to Google's organizational structure and the difficulty of shipping new products (although he acknowledges that the early models were rather incoherent, and the story should be read with that qualification).
He joined OpenAI in 2024, seeking closer collaboration between research and product. He started with infrastructure again, then worked with colleagues to train models on the internal Python codebase and build agents to accelerate research. Those experiments merged with the Autonomous Software Engineer effort and became one of the foundations of Codex. Tibo also acknowledges that the early cloud version did not find product-market fit: there was too much friction for users. Being useful internally does not automatically make something convenient externally.
The conversation then offers several details about how this changes development itself:
🔸 The agent needs architecture too The Codex core was separated from its interfaces and written in Rust for reliability, security, and efficiency, even though models were better at Python and TypeScript at the time. The team was already thinking about running the agent across products and at scale. 🔸 Some scaffolding should be able to die off The harness, the environment providing tools and instructions, compensates for model weaknesses. Initially, a model needs reminders to run tests; later, the behavior becomes part of the model itself. Tibo describes engineers and researchers deciding together where to fix a problem and how long to wait for the next model. You might build an elaborate workaround that becomes unnecessary a month later. 🔸 Code review is shifting toward intent and contracts According to Tibo, OpenAI is automating correctness and security checks, with flagged security issues blocking pull requests. He suggests focusing human discussion on what a component should do, which data it may access, and which invariants it must preserve. Those agreements are useful before generating the implementation. 🔸 The agent needs the history of decisions Internally, Codex is connected to code, Slack, and documents. During the merger of Codex and ChatGPT, it even chronicled discussions and decisions. That is an interesting use for an agent: helping a team reconstruct why a system ended up the way it did.
This continues the theme of AI for software architecture: boundaries, constraints, and the history of trade-offs need to be available to work with. Alongside it remains the question from Geoffrey Litt's talk about understanding: how much of this can the team itself explain?
In Tibo's assessment, maintenance and architectural changes are becoming much cheaper. But his reasoning contains a condition: good abstractions let you change a component's internals without affecting everything around it.
#AI4SDLC #AI #Agents #Architecture #Engineering