Governance at the Boundary: How Agent Decomposition Degrades Policy Compliance
2026-08-17 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors created Fiducia-bench, a test to see if financial AI agents follow rules correctly, like escalating issues or leaving records. They studied whether breaking an agent into smaller parts makes it worse at following these rules. They found that when information passes between parts, important facts often get lost, causing mistakes. This problem was worse with less powerful models and appeared as both missing risks and missing clearances. The authors also shared all their tasks and tools publicly.
agent benchmarksfinancial agentsgovernancepolicy adherenceKYCAMLmodel decompositioninformation attenuationescalationaudit trail
Authors
Bowen Li, Guojun Wang
Abstract
Existing agent benchmarks ask whether the agent finished the task. We ask whether it finished it within policy. We introduce Fiducia-bench, a benchmark for the governability of financial agents---whether they escalate when obligated, abstain when required, and leave an auditable trail---and use it to study a question no prior benchmark addresses: does decomposing an agent into components degrade its governance? It does, and the mechanism is specific. Policy-relevant facts discovered by one component are attenuated at the handoff boundary before reaching the component that must act on them. In a 626-episode experiment across 100 KYC/AML task variants, two models, and three architectures, a 32B open-weights model attenuated 0% of discovered facts under a single-loop baseline, 56% under a fixed pipeline, and 85% under an orchestrator-subagent architecture (all at constraint distance 2). A stronger model (gpt-4.1-mini) attenuated 3-6% under the same conditions, suggesting the governance cost of decomposition is partly a function of model capability. Critically, the same mechanism produces both under-escalation and over-escalation, depending on whether the dropped fact was a risk signal or an exculpating one. The benchmark, all tasks, and the verification harness are open-source