Watching the OpenAI Dev Day videos, I listened as Thibault, engineering lead for Codex, announced “Codex is now a senior engineer.”
AI entered the organization as an intern - uncertain & inexperienced. Over the summer, engineering leaders said treat it like a junior engineer.
Congratulations, Robot. You’ve been promoted - again! From intern to senior engineer in about a year. Quite the trajectory.
Other data points :
- 92% of technical staff use Codex daily
- those staff generate 72% more pull requests (code submissions) than those who don’t use AI
The team shared more. The best design patterns for collaborating with Codex are architect-implementer systems & closed feedback loops.
ARCHITECT-IMPLEMENTER
I wrote about architect-implementer architectures on Monday. The pattern splits work between two separate robots : the first designs the solution, the second executes it.
Ask a robot to write the plan document. You’ll refine your thinking as you review it. The robot manages progress through each step.
The counterintuitive part? The second robot shouldn’t see the first robot’s context. Fresh discerning digital eyes catch more errors.
CLOSED FEEDBACK LOOPS
In the plan, designing the tests / hurdles that a robot must pass to complete the task is critical. The robot runs the tests, fixes the code, runs the tests again, and repeats until passing. These tests can be visual (evaluate screenshots), functional (does the code run), or logical (does the code meet the requirements). Then a third robot reviews for quality & style.
The record at OpenAI is 7 hours of autonomous execution, 150M tokens, and 15K lines of code refactored with this design pattern. Pretty remarkable even for a senior engineer.
Congratulations, Robot. Keep climbing that ladder.