When AI agents team up, is a permanent shared record what makes them work well, or something else?

This explores how much of a team's coordination benefit comes from a shared, permanent record versus everything else that's going on. The corpus doesn't contain a clean measurement. The closest case is a group of thirteen language-model workers with no central planner. They coordinated only through an append-only Git record of results and lineage, and over 12 days they produced 1,703 contributions and closed 62% of the gap to a trained baseline Can decentralized agents coordinate research without a central planner?. As summarized, there is no run without the record to compare against. The 62% is the whole package, not the record's share of it.

Nearby notes bound the question from different sides. MetaGPT found that agents passing standardized documents coordinate better than agents chatting Does structured artifact sharing outperform conversational coordination?. That result also bundles the document format, agents pulling what they need instead of being broadcast to, and human-style work procedures. So it shows that the form of the record matters, but it doesn't say how much. In a single-agent version of the same idea, a 20B search model with its bookkeeping moved into an external harness beat the next open searcher by 11.4 points and matched frontier models Can externalized bookkeeping let smaller search agents beat larger ones?. The gains carried over to held-out benchmarks, which the note reads as the harness being a learned capability. Part of a record's benefit may therefore come from the agent having learned to use it well, not from the record alone.

Three other notes point to benefits a record can't supply. First, a check that only reads the transcript can't reliably tell whether a reflection helped when the truth depends on the outside world, while a check grounded in the environment can Can transcript alone tell whether a reflection helps?. My inference is that a record pays off most when it stores checkable results, as the Git record does, rather than talk about results. Second, a batch of diverse, mediocre contributions can beat redundant expert ones, because whoever combines them needs raw material to arbitrate between Can diverse mediocre traces outperform redundant expert traces?. Some of a decentralized team's gain may come from its contributions being varied, with the record only keeping them from being lost. Third, in a 535-person study, people assisted by an LLM captured only about half of the item-level accuracy gain, so making complementary strengths available doesn't guarantee they get used Why does assisted accuracy capture only half the LLM gain?.

The collection supports a qualified answer: the record looks like necessary infrastructure, not the whole engine, but no note puts a number on its share. The experiment that would is the same agents on the same task under three conditions: with the record, with chat only, and with the record but redundant contributors. A chain-of-thought study shows what that kind of attribution looks like. It split performance into output probability, memorization, and noisy reasoning, and output probability alone swung accuracy from 26% to 70% What three separate factors drive chain-of-thought performance?. The collection has no equivalent split for coordination records.