Building Analyst-Like Agents: A Self-Improving Multi-Agent Framework for Financial Reasoning in the Enterprise

Enterprise accounting data is complex, ambiguous, and shaped by evolving systems and regulations. The institutional knowledge needed to reason over the data is sparse, scattered and rarely structurally documented—posing major challenges for LLM agents. We introduce a multi-agent financial research framework that mimics a junior analyst’s onboarding and growth. The Analyst Agent learns proactively from repeated month-end cycles, builds long-term memory, clarifies ambiguity with an Accountant Agent, and collaborates with an Engineer Agent to refine tools when needed. This self-learning, self-reflecting, and tool-refining workflow enables the agent to adapt to vague conventions, reason with "business sense", and validate its own analysis. Evaluated on 200 realistic accounting questions across four month-end cycles, our system boosts first-response accuracy from 44.5% to 81.3%, with measurable gains in reasoning, clarification, and tool use efficiency. Our agent learns like humans, grow like humans, and ultimately reason like humans—while working in a enterprise world that is messy, ambiguous, and alive.