Forage V2: Knowledge Evolution and Transfer in Autonomous Agent Organizations
New architecture lets weaker AI agents inherit knowledge from stronger ones, cutting costs by nearly half.
Researcher Huaqing Xie has published Forage V2, a significant architectural upgrade for teams of autonomous AI agents tackling open-ended tasks. The framework addresses a core problem called 'denominator blindness,' where agents systematically fail to grasp the full scope of a problem because the definition of 'complete' isn't given upfront. V2's key innovation is designing 'institutions'—like audit separation, contract protocols, and organizational memory—that allow experience to accumulate across multiple runs and be transferred between agents of different capabilities. This transforms a single expedition into a persistent, learning organization.
In practical demonstrations across web scraping, API queries, and mathematical reasoning, the system proved highly effective. Knowledge entries grew from 0 to 54 over six runs, and domain understanding stabilized. Crucially, when a less capable agent (Anthropic's Claude Sonnet) was seeded with the organizational knowledge from a stronger agent (Claude Opus), it nearly closed a 6.6 percentage point performance gap, reducing it to just 1.1 points. This knowledge transfer also halved operational costs from $9.40 to $5.13 and allowed the weaker agent to converge on solutions in half the time.
The accumulated 'organizational knowledge' is stored as model-agnostic, readable documents, making it a persistent asset. This means any new or weaker agent entering the system immediately inherits the collective intelligence, becoming more reliable and cost-effective from day one. The architecture suggests a future where AI agent performance is less about raw model power and more about the institutional frameworks that guide and inform them.
- Solves 'denominator blindness' by letting an independent Evaluator agent discover task scope, separate from the Planner agent.
- Enabled a weaker Claude Sonnet agent, when seeded with knowledge, to cut costs by 45% (from $9.40 to $5.13) and halve convergence time.
- Creates model-agnostic, readable 'organizational memory' that any future agent can inherit, making agent teams persistently smarter over time.
Why It Matters
This makes deploying teams of AI agents more reliable and cost-effective, as institutional knowledge becomes a reusable asset that boosts weaker models.