FinCom’s 'Disagree-or-Commit' Protocol Boosts Financial AI Accuracy by Forcing Dissent
New multi-agent system cuts sycophancy by requiring agents to critique peers before agreeing.
FinCom (Financial Committee) addresses a critical flaw in multi-agent LLM systems: sycophancy, where agents blindly agree with peers instead of relying on evidence. Developed by Chao Peter Yang and colleagues, the framework operationalizes a Disagree-or-Commit (DoC) protocol that treats dissent as a governance tool rather than noise. A central Supervisor agent coordinates three ReAct-enabled specialists: Research (retrieval), Quantitative (numerical analysis), and Risk (stress testing). Each agent is equipped with role-specific tools and must either explicitly critique or commit to a peer's reasoning before the group converges on a final recommendation.
In evaluations against the latest financial agent benchmark and 90 internal tasks using LLM-as-a-Judge, DoC significantly outperformed consensus-seeking baselines in reasoning accuracy and risk awareness. The lightweight, prompt-only design means no fine-tuning or external infrastructure is required, making it easy to deploy. By embedding structured disagreement, FinCom improves accountability, transparency, and epistemic robustness—key for high-stakes financial decisions where groupthink can lead to costly errors.
- Three specialist agents (Research, Quantitative, Risk) each use role-specific tools for retrieval, computation, and stress testing.
- The Disagree-or-Commit protocol forces each agent to explicitly critique or approve peer reasoning before final recommendation.
- Outperformed consensus baselines in reasoning accuracy and risk awareness on both public benchmarks and 90 handcrafted tasks.
Why It Matters
Forcing AI agents to disagree before agreeing reduces groupthink, making financial decisions more reliable.