Forage V2: Knowledge Accumulation and Transfer in Autonomous Agents

A recent paper on arXiv (2604.19837) presents Forage V2, an upgraded version of the Forage framework designed for autonomous agents engaged in open-world tasks. This iteration tackles the issue of denominator blindness, where agents consistently underestimate the extent of their target space, by refining evaluation techniques and isolation strategies. Forage V2 facilitates the accumulation of knowledge across multiple runs and allows for the transfer of insights between different model capabilities, incorporating safeguards to prevent degradation. In experiments involving web scraping, API queries, and mathematical reasoning, knowledge entries increased from 0 to 54 over six iterations, while a less capable agent (Sonnet) improved its performance by leveraging knowledge from a more advanced agent (Opus).