Lena Osei-Mensah

Safely retrying agent actions requires idempotent tool design across three layers.

Fixing the data fetching layer reveals why most autonomous research agents fail on the open web.

Controlling what a research agent crawls determines whether its answers are trustworthy or noise.

Dividing work by cognitive function beats generalist models on latency, isolation, and scale.