Should AI agents coordinate by chatting with each other, or by passing along finished work with a traceable trail?

This explores whether keeping a traceable trail of how work was produced, with each agent's output passed forward intact, coordinates agents better than having them talk to each other. The corpus has no note on 'lineage preservation' by name, so this answer reads it through the nearest evidence. That evidence leans toward intact work products over conversation, and the reasons are more interesting than format alone.

The most direct comparison is MetaGPT. Its agents hand each other standardized engineering documents and pull what they need from a shared environment, and this beats chatting back and forth. Pulling from a shared record also removes conversational noise, much as human teams lean on specs and tickets rather than hallway talk Does structured artifact sharing outperform conversational coordination?. Each document is a checkpoint that can be traced back to whoever produced it. That trail comes free with the structure, and a chat log doesn't give you it.

Conversation's weakness shows up when it scales. On the AgentsNet benchmark, agents coordinate worse as the network grows. They agree too late, or adopt a strategy without telling their neighbors. They also accept neighbor information without checking it, so errors spread, even though they can still spot direct conflicts Why do multi-agent systems fail to coordinate at scale?. The corpus doesn't test provenance directly, but the failure fits: a claim in a chat message arrives without its origin, so there is nothing to check it against. Multi-agent deliberation also has its own failure modes, such as Degeneration-of-Thought and Silent Agreement, which suggests that more talking does not reliably mean better reasoning What limits autonomous capability in large language models?. A trail matters most for the reason single agents get replaced at all. Tasks that need independent verification and parallel specialists overflow one agent's loop Do single agents always hit organizational limits?, and verifying someone's work requires seeing how it was made.

There is also a third option, and it changes what the comparison is about. Both lineage and latent sharing try to stop information degrading when reasoning is squeezed into natural language. LatentMAS lets agents pass internal representations directly through KV caches. It gains 14.6% in accuracy and uses 70.8–83.7% fewer tokens with no extra training, because hidden embeddings keep fidelity that text loses Can agents share thoughts without converting them to text?. A related approach uses sparse autoencoders to separate each agent's private thoughts from shared ones, so alignment conflicts show up before they surface in language Can agents share thoughts directly without using language?. Lineage preserves the trail of finished artifacts, while latent sharing preserves the model's internal state. Chat is the lossy middle layer that both bypass.

Keeping everything has a cost, and the memory notes show how agents deal with it. DeepAgent folds its interaction history into episodic, working, and tool memory schemas. That saves tokens and lets it pause and rethink, provided the agent does the folding itself and the structure is sound Can agents compress their own memory without losing critical details?. MRAgent goes the other way and rebuilds memory on demand by walking a graph and pruning paths as evidence builds. It gains up to 23% on reasoning tasks and costs less than fixed retrieval Can agents reconstruct memory on demand instead of retrieving it?. Together they suggest a practical answer: keep the lineage, but store it compactly enough to walk when needed.

The choice doesn't have to be either/or. Coordination standards tend to win by wrapping existing protocols rather than replacing them Should coordination protocols wrap existing systems or replace them?. A lineage layer would probably spread the same way, by sitting underneath conversation rather than banning it. The corpus does not say how much lineage helps beyond structured artifacts, and that is the open gap.