Field notes

The lab notebook, published.

What we learn building agentic systems, written down while it's still sharp. Eval design, failure taxonomies, tooling — and the occasional detour into the weeds.

June 12, 20266 min read
The Verifier's Dividend
A checker that catches 90% of an agent's mistakes buys the same task horizon as ten months of frontier model progress — for under 1% extra compute. Four instruments and one falsifiable claim.
agentsreliabilityverification

The lab notebook, published.

The Verifier's Dividend