Splitting an AI agent into specialized components makes it worse at following compliance rules, according to new research.
Researchers built Fiducia-bench, a benchmark that tests whether financial AI agents escalate risks, abstain from acting, and leave an audit trail when required by policy, rather than just checking if they complete a task. Across 626 episodes covering 100 KYC/AML task variants, two models, and three system architectures, the team found that facts relevant to policy decisions get lost when they cross the handoff boundary between agent components. A 32B open-weights model dropped 0% of discovered facts in a single-loop setup, 56% in a fixed pipeline, and 85% in an orchestrator-subagent architecture. A stronger model, gpt-4.1-mini, dropped only 3-6% of facts under the same conditions.
This matters because orchestrator-subagent architectures are becoming the default way companies build complex AI agents, on the assumption that breaking work into specialized pieces is strictly an engineering upgrade. The research shows that decomposition carries a governance cost, and that cost cuts both ways: dropped facts cause agents to both under-escalate real risks and over-escalate on cases that should have been cleared.
In other words, the more modular your agent, the more you need to check that policy-relevant facts survive the handoff - a problem no amount of prompt engineering on the final component alone will fix.