Technical interest · T-10
Agent evals
Measuring what AI agents actually do: benchmarks, harnesses and scoring that hold up when you look closely.
Agent Ledger benchmarked web agents on 24 tasks across 10 sites with WebArena-verified scoring and Wilson lower bounds, and found that naive memory retrieval cut success from 21% to 13%. The AI Safety & Agent Harness fellowship gates every generated specification through a five-gate validation oracle, and the inter-agent misalignment project is built on a two-agent harness.
Nearest neighbors · any layercosine distance in the full space