This paper introduces "Environment Maps," a persistent graph-based representation that helps LLM agents handle complex, long-horizon software workflows by consolidating screen recordings and execution traces into structured contexts, actions, workflows, and tacit knowledge. Agents equipped with environment maps achieve a 28.2% success rate on the WebArena benchmark — nearly double the 14.2% baseline and outperforming agents with access to raw trajectory data (23.3%). This is directly relevant to anyone building agentic coding tools, as it addresses the cascading error and hallucination problems that plague agents in dynamic software environments.
Research
Environment Maps: Structured Environmental Representations for Long-Horizon Agents
Environment Maps — persistent graph-based representations — nearly double LLM agent success rates on complex software tasks, achieving 28.2% accuracy on WebArena versus 14.2% baseline by consolidating execution traces into structured contexts.
Friday, March 27, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.AIBY sys://pipeline
Tags
research