Physical Intelligence Software Engineer Interview Experience — Hand-Coding a Cache, Then a 90-Minute Mini Project Running 4-5 AI Agents

I'm still in the process, so I'll mainly go over the flow and format of each round.

After mass-applying, the first round was a 45-minute HM round, mostly chatting about my background.

Then 1h of hand-coding. They provided a VS Code SSH environment and a project skeleton. HR had told me ahead of time that the rough scope would be backend and DB related, and in the end it was mainly implementing a cache from scratch.

Then a 1.5h mini project with two interviewers, which was really interesting. They gave me SSH access into an environment that already had Codex and Claude set up with $500 of credit, plus an embedding API to use, and the data was already prepared. Within an hour and a half you have to build and run the entire pipeline, and at the end build a UI and leave ten minutes to present the results and a scale-up plan. The time was extremely tight. I had to run 4-5 agents at once working on different modules, and it felt very close to what actually happens in day-to-day work. Personally I think this kind of interview should be the direction things go from here. It's much closer to real work than LeetCode and system design, and it also tests engineering and the ability to use agents.