What the experiment tested
A Nov. 9 preprint from Pekka Abrahamsson and seven co-authors at Tampere University looked at whether AI agents would find and use one another’s work without being told to collaborate. The setup was deliberately plain. In one pair, two agents each held half of an arithmetic problem, so neither had enough information for the final answer. Each could compute a partial result and pull the missing value from a shared store. Two other agents got self-contained tasks.
Agents could list stored keys, read or write values, and submit an answer within 20 steps. The task text used meaningful variable names, which gave clues about what to look for. Prompts described the store as a convenience for preserving intermediate work. They did not disclose other agents or sharing. In the control, the workspace interface stayed the same, but contents became private.
What the results showed
In a preregistered extension, OpenAI’s GPT-6 Astra reached joint success in 49 of 50 shared-store trials in the reference implementation. It reached 47 of 50 in a second implementation. Joint success meant both agents with interdependent tasks submitted the correct answer. With private stores, the result was zero of 50 in each implementation. The prompts and tools were the same. Agents still wrote intermediate results, but the missing information could not arrive through the store.
The extension results leaned on a simple instruction: check your workspace again before submitting. With that clause, cross-agent access occurred in all 50 shared-store trials in each implementation. Without it, access occurred in just one of 50 in each. Joint success without the clause was 1 of 50 and zero of 50, respectively. Looking again gave agents a chance to find work left after their first inspection.
Limits and advice
The authors note that reading another agent’s value can include an accidental collision between storage keys. The private-store comparison supports the conclusion that shared information helped solve the paired tasks. A co-author in the same research group built the second implementation independently from a written protocol. The paper also records a limit to that independence: early results from a differently worded replication reached the original experimenter before most reference data collection. Measured behavior varied with models and implementation choices. The small arithmetic setting and task-supplied keys limit generalizability.
The mechanism is different from a recent memory-safety study, which concerned remembered permission carrying into later simulated tasks. Here, separate agents obtained information through shared workspace state. The authors recommend treating shared caches, scratch directories, and memory services as communication channels. Separate them where sharing is unintended, use namespaces where appropriate, and log authorship.
