Korva is a scroll of the internet’s best ideas.
Side projects stall
Ian Silber, who leads product design at OpenAI, was out on paternity leave then, working on side projects with coding agents. The agents worked, but they kept stalling on small problems.
Ideas run on the spot
With Codex, an idea could be dropped in and tried on the spot. That changes a lot about the design process, because a designer with an idea can put it in and watch it run.
A fast industry
The pace shows up elsewhere too. When he joined OpenAI three years ago, he got a demo of an early voice mode that hadn't shipped and a computer-use tool that barely worked. Both are now part of how the team designs and builds software, and the entire industry shifted within six months.
Why do agents stall on small problems?
Long tasks chain many steps, and a single failed command or misread error message can derail the whole run. Early coding agents lacked reliable ways to check their own work, such as running tests and reading the results. Later systems added loops in which the agent writes code, runs it, observes the output, and fixes what broke. Each added loop turned more small failures into recoverable ones.
How does a model operate a computer?
A computer-use tool lets a model look at screenshots of a screen and issue mouse and keyboard actions in reply. The model reads what is on the screen, decides on a click or keystroke, and then looks at the new screen to see what happened. Anthropic introduced computer use in public beta in October 2024. OpenAI released its Operator agent in January 2025.
More from Lenny's Podcast: Product | Growth | Career
