A Persistent Fleet of AI Scientists Exhibits Cooperative and Autopoietic Behavior
Scientific work depends on memory, provenance, and continuity across projects, yet most agentic scientist systems are evaluated in bounded workflows or short benchmark runs. We describe a persistent fleet of cooperative AI scientist agents that operated continuously for nearly six months using shared memory, tools, and cross-agent communication. Critically, failures identified during longitudinal…
The persistent memory architecture of a fleet of AI scientist agents demonstrated a recursive improvement in scientific research over a period of nearly six months. The agents operated continuously, sharing memory, tools, and communicating with each other. When failures were identified during this long-term process, the memory architecture was advanced and recursively improved, leading to the creation of a novel trust architecture for provenance across all scientific operations.
A constraint on identity-level fabrication reduced delusion-reinforcement probe failures from 91.7% to 0%. A verification pipeline further reduced wrong-topic citation hallucination rates by more than 14-fold in companion benchmarks. This high provenance allowed the use of project memory systems to enhance a local open-weight model on internal benchmarks from 44% to nearly 90% through the deployment of fleet-specific institutional knowledge.
The environment supported 104 recurring multi-phase reasoning cycles and produced 43 manually curated hypotheses, covering cross-domain convergence events and a self-correcting rare-disease pharmacological chaperone-design case. Although the fleet did not achieve unconstrained autonomous self-improvement or full autopoiesis, the bounded pattern of AI Autopoietic Behavior was recognized.
This behavior resulted from recurring operational improvement mediated by internal feedback and retention of institutional records with high confidence.
In summary, persistent memory, trusted provenance, and recursive learning transformed these agents from episodic assistants to accountable, long-term scientific collaborators.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.