Researchers have introduced MoCA-Agent, a code agent for financial and numerical reasoning built around a market of claims.
Instead of relying on free-form debate, the system decomposes a question into typed atomic claims and has specialist agents buy or sell those claims. Accepted claims are then used to synthesize executable Python code.
The approach targets a common weakness in financial QA: a single misread cell, unit, or formula can produce a fluent but wrong answer.