Forage V2: Knowledge Evolution and Transfer in Autonomous Agent Organizations

2026-04-01 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Forage V2 introduces an architectural framework for autonomous agent organizations to accumulate and transfer knowledge across runs, addressing "denominator blindness" in open-world tasks where completion boundaries are unknown. Building on V1's co-evolving evaluation and method isolation, V2 enables agents to extract and store lessons in a persistent, append-only knowledge base. Experiments across web scraping (NVIDIA GPU collection), API queries (UniProt proteins), and mathematical reasoning (First Proof Q10) demonstrate that knowledge entries grow from 0 to 54 over six runs, stabilizing denominator estimates. Knowledge transfer allows a weaker agent (Sonnet), seeded with a stronger agent's (Opus) experience, to narrow a 6.6 percentage-point coverage gap to 1.1pp, halve costs from $9.40 to $5.13, and converge in half the rounds (4.5 vs. 7.0). The system's design emphasizes institutional safeguards like physical workspace isolation and advisory knowledge delivery.

Key takeaway

For AI Architects designing autonomous agent systems for open-world tasks, you should prioritize architectural integrity over individual agent capability. Implement robust method isolation and an organizational memory system like Forage V2's knowledge base. This approach ensures credible evaluation and enables weaker models to achieve strong performance by leveraging accumulated institutional experience, significantly reducing operational costs and improving convergence speed.

Key insights

Institutional design, not individual agent enhancement, makes autonomous agents reliable in open-world tasks.

Principles

• Method isolation prevents self-serving evaluation.
• Knowledge accumulation improves efficiency and calibration.
• Advisory knowledge guides without prescribing.

Method

Agents independently extract lessons post-mortem, storing them in an append-only knowledge base. This organizational memory is delivered via system prompts, allowing subsequent runs and weaker agents to inherit accumulated experience.

In practice

• Implement physical workspace separation for agent isolation.
• Use append-only knowledge bases to preserve historical observations.
• Design evaluation criteria to co-evolve with task execution.

Topics

• Forage V2 Architecture
• Autonomous Agent Organizations
• Knowledge Evolution
• Knowledge Transfer
• Denominator Blindness

Code references

• openai/symphony
• karpathy/autoresearch

Best for: AI Architect, Research Scientist, AI Scientist, AI Engineer, MLOps Engineer

Related on AIssential

• Building Self-Healing Software Systems with Multi-Agent AI Architectures — Multi-agent AI enables autonomous detection and resolution of software system issues.
• Capability Architecture for AI-Native Engineering — A public reference model is crucial for coordinating AI integration and establishing shared engineering standards.
• Agents are ready but your architecture probably isn't — AI value stems from outcome-driven design, robust data infrastructure, and proactive governance, not just model selection.
• Knowledge Activation: AI Skills as the Institutional Knowledge Primitive for Agentic Software Development — Knowledge architecture, not model capability, is the bottleneck for effective agentic software development in enterprises.
• SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems — Socialized evolution enables breakthroughs for plateaued agents, but gains are conditional on abstraction and context.
• Conversational ML Workflows via the H2O AI Super Agent in Enterprise h2oGPTe | Part 21 — The h2O Super Agent enables autonomous system orchestration via natural language, acting as an IDE.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.