Talks¶
Better Harness, Better Answers¶
Document intelligence without bigger models: one small model stays fixed, the harness changes one step at a time, and every score comes from logged runs on real public documents.
Open the deck · Code · Dataset
Use the arrow keys or space to move between slides. Press n for speaker notes and f for full screen.
The Agent as Compiler¶
Harness engineering: why coding agents (and LLM products) are best understood as compilers of human input, the loops that run a harness, and why the reward model and environment are what make the model interchangeable.
Use the arrow keys or space to move between slides.