Ben Stolovitz now manages five or six coding agents and says machines write over 90% of his code. His annual review describes enormous gains in building software, alongside exhaustion from supervising it. He still wants human judgment to guide architecture, research, and creative work.

Coding gets faster; checking still takes work

Stolovitz moved from Microsoft’s Windows AI team to Salesforce AI Research. He reports spending about $3,000 a month on tokens while supporting scientists and building prototypes.

  • Give each agent a separate workspace. He uses Git worktrees, independent copies of a repository, so agents can implement features, reproduce bugs, run tests, and prepare draft pull requests.
  • Review the design and test the result. He focuses on architecture and recurring problems, then turns those problems into review rules and better automated tests. Human validation remains part of the process.
  • Keep simplification on the table. His most successful work this year was a hand-written tool that replaced roughly 4,000 generated lines with a few hundred.
  • Budget attention as well as tokens. Six simultaneous conversations make him feel overstimulated. The temptation to send one more prompt keeps intruding on breaks.

Finding information differs from understanding it

AI has become his first choice for search and finding research papers. It also helps him practice established material, including quizzes about code an agent wrote. But tutors can slip back into giving answers, removing the effort that learning requires.

His experience designing a voice turn-detection model was worse. Agents helped him explore a design without suggesting better architectures until he explicitly asked. Days of confusion about labeling training data ended after one discussion with a human teammate. Fluent assistance can keep a weak approach moving.

Decide which parts of making things to keep

Local dictation with OpenWhispr and Handy is already useful to him. Open-ended voice conversation still feels awkward. He rejects AI-written personal prose and finds editing advice bland, although agents sometimes document code and experiments they completed themselves.

He also distinguishes creative play from commercial content flooding. His proposed personal AI policy starts with wanting to make things with other people. That gives him a reason to preserve programming, writing, drawing, and music even when generation becomes cheaper.

Caveat: This is one well-funded research engineer’s experience. His prototype speeds and productivity claims are personal reports, not controlled comparisons.