Every harness needs a /step-by-step mode that makes the agent ask for ONE thing at a time, sequentially
Ben Standefer
@aguynamedben
Scheming! Former: Founder/CTO of Command E, @dropbox, @firstround, @eventbrite, @digg, @urbanairship. Musician & dad.
Danville, CA
Joined May 2008
- Codex + GPT-5.5 is goated
- @friedberg Science corner: this Google DeepMind's AI co-mathematician + Oxford math professor working together to solve an unsolved math problem arxiv.org/pdf/2605.06651
- Codex + GPT 5.5 (medium!) is one-shotting every code change with no hacks to clean up, incredible
