Skill-centric training beats task-centric for small web agents across 4 benchmarks
anmarasovic · x · 2026-11-03
New research finds small web agents should practice skills rather than whole tasks: with the same amount of data, a 4B agent trained on skill-centric traces beats a task-centric counterpart on all four benchmarks tested. Paper and thread linked.
More from coding & agent
- Pinokio launches Remote feature for one-click open-source AI tool installs — cocktailpeanut · 2026-11-03
- Runway launches Runway Dev, an API platform built for coding agents — runwayml · 2026-11-03
- The approval queue is not the tax — it's where the agent learns — AIGODSEND · 2026-11-03
- Proxy vs tool-call gate: where should agent guardrails actually sit? — quiet_flare · 2026-11-03
- Solo dev builds nationwide 3D camper map with 15k bridges modeled to spec — tomjohndesign · 2026-11-03
- Agent harness debate: is the opinionated layer above the core harness a new thing? — sull · 2026-11-03