Forage V2: Knowledge Evolution and Transfer in Autonomous Agent Organizations
Summary: arXiv:2604.19837v1 Announce Type: new
Abstract: Autonomous agents operating in open-world tasks — where the completion boundary is not given in advance — face denominator blindness: they systematically underestimate the scope of the target space. Forage V1 addressed this through co-evolving evaluation (an independent Evaluator discovers what “complete” means) and method isolation (Evaluator and Planner cannot see each other’s code). V2 extends the architecture from a single expedition to a learning organization: experience accumulates across runs, transfers across model capabilities, and institutional safeguards prevent knowledge degradation.
Key Findings
In the latest development of Forage V2, researchers demonstrate significant advancements in two main areas: knowledge accumulation and knowledge transfer. These findings were validated across three task types, including web scraping, API queries, and mathematical reasoning.
Knowledge Accumulation
The study reveals that over six runs, knowledge entries grew from 0 to 54, showcasing a significant increase in information retention. This growth is accompanied by stabilizing denominator estimates, indicating a deepening understanding of the domain. The implications of this knowledge accumulation are profound, as they suggest that autonomous agents can progressively refine their understanding and improve their performance over time.
Knowledge Transfer
One of the most compelling aspects of Forage V2 is its ability to facilitate knowledge transfer. The research highlights the performance of a weaker agent, referred to as Sonnet, which was seeded with the knowledge of a stronger agent, Opus. The results of this knowledge transfer were remarkable:
- The coverage gap narrowed from 6.6 percentage points to just 1.1 percentage points.
- The operational cost was halved, decreasing from $9.40 to $5.13.
- The convergence time was significantly reduced, with Sonnet converging in an average of 4.5 rounds compared to Opus’s 7.0 rounds.
Furthermore, three independent seeded runs successfully arrived at the same denominator estimate of 266, demonstrating that the organizational knowledge effectively calibrates evaluation processes.
Architectural Contributions
The contributions of Forage V2 extend beyond empirical results; they introduce architectural innovations that enhance the reliability of autonomous agents. Key institutional designs include:
- Audit Separation: Ensures that evaluation and planning components remain independent to reduce biases.
- Contract Protocols: Establish clear agreements and responsibilities that guide agent interactions.
- Organizational Memory: Captures and stores accumulated experiences in a model-agnostic manner, allowing future agents to inherit valuable insights regardless of their capabilities.
These innovations not only enhance the reliability of any agent entering the organization but also ensure that the accumulated experience is accessible and beneficial for future developments.
Conclusion
Forage V2 represents a significant advancement in the field of autonomous agents, offering a robust framework for knowledge evolution and transfer. By building institutions that support learning and collaboration, Forage V2 sets the stage for more effective and efficient autonomous systems in open-world environments.