Starfish-FL: Harnessing Agentic Federated Analytics

Federated learning (FL) enables privacy-preserving multi-site clinical research, but orchestrating, monitoring, and interpreting federated analyses remains out of reach for the biostatisticians and clinicians who most need them. We present Starfish-FL, a multi-tier agent harness for cross-silo healthcare analytics built around the safety invariant that the LLM is advisory and cannot independently trigger consequential actions. Agents operate at three tiers. An Autonomous Experiment Agent orchestrates end-to-end experiments via CLI tools, an Embedded Router Agent handles adaptive aggregation, early stopping, and failure triage, and Controller Agent Hooks produce per-site summaries, outlier flags, and convergence signals. Three hybrid safety policies, SafeStop, SafeAggregate, and SafeTriage, gate every site exclusion, early stop, and recovery action behind deterministic preconditions, so LLM errors cannot cause unsafe state transitions, conditional on guard correctness. All agent features are opt-in and degrade gracefully to standard FL, and the platform provides 21 Python and R task implementations with built-in statistical diagnostics. Evaluated across four LLM providers with repeated-run confidence intervals, task selection reaches \({\sim}97\%\) accuracy with hosted models, the online SafeStop policy saves 50–70% of rounds with no accuracy loss, and safety-policy validation shows a 100% safe-action rate under injected failures, confirming that deterministic guards, not the LLM, preserve safety.