Tag
This paper formalizes task-agnostic environment preprocessing for LLM agents, where an agent explores an unfamiliar environment to produce reusable artifacts for a frozen solver, with meta-agent variants achieving high performance across benchmarks.
A Twitter thread highlights the limitation of AI agents where useful runs die with the session, and proposes the idea of turning AI workflows into reusable, memory-enabled artifacts that can be deployed as desktop apps without consuming tokens.
Introduces SKILL.nb, a framework for governing reusable agent workflows through evidence-calibrated lifecycle policies, featuring selective formalization and gate-conditioned execution. It achieves significant improvements on web automation benchmarks and demonstrates resilience to environment drift.