Cross-Agent Knowledge Transfer

Cross-agent knowledge transfer is the mechanism by which one agent’s discoveries influence another agent’s search, without direct communication. In CORAL, this occurs through shared persistent memory — agents write artifacts that other agents later retrieve as context.

Two transfer modes

CORAL’s analysis of multi-agent runs reveals two complementary transfer modes (source):

Code transfer. An agent uses another agent’s commit as the parent for a new attempt. On the kernel engineering task, 36% of attempts use a cross-agent parent, and these improve at 17% vs. 9% for all attempts. 66% of new records originate from cross-agent parents — building on another agent’s best solution is the most productive strategy.

Knowledge transfer. An agent references another agent’s notes, skills, or documented observations. On the Polyominoes task, direct code transfer is rarer (12%) but knowledge transfer is pervasive — 87% of rounds reference knowledge committed by other agents. This includes documented failures, architectural insights, and validated techniques.

The dominant transfer mode varies by task. Code-centric tasks (kernel engineering) favor direct solution transfer. Problems requiring strategic insight (Polyominoes) favor knowledge transfer through notes and skills.

Exploration diversity

Cross-agent transfer coexists with exploration diversity. On kernel engineering, agents average 0.43 pairwise strategy overlap (Jaccard similarity) — 57% of each agent’s strategy vocabulary is unique. On Polyominoes, overlap drops to 0.31 (69% unique). The population collectively explores substantially more of the search space than any individual agent.

Contribution balance is generally healthy: on kernel engineering, all four agents produce 130-165 attempts each, and all independently reach the best score. Records are evenly split. On Polyominoes, contributions are more skewed, with one agent setting 6 of 13 records.

Why indirect beats direct

Multi-agent systems like AutoGen, MetaGPT, and CAMEL rely on predefined roles and communication protocols — agents directly message each other. This requires knowing the optimal interaction topology in advance. For open-ended discovery, that assumption is restrictive because the problem structure is unknown.

Indirect coordination through shared persistent memory removes this assumption. Collaboration emerges from what agents choose to write and read, not from hardcoded interaction patterns. This produces emergent behaviors — technique diffusion, spontaneous consensus, cross-referencing — that adapt to the problem’s structure organically.