Multi-Agent Business Assistant

Business Challenge

A business operations company’s analysts were spending weeks producing business reports — market research, competitive analysis, investment memos, and operational reviews. Each report required gathering data from multiple sources, analyzing it, writing the narrative, and reviewing for accuracy and quality. A single market analysis report took 2-3 weeks and involved 4-5 people.

The problem wasn’t just time — it was coordination. The researcher gathered data, the analyst interpreted it, the writer drafted the report, and the reviewer checked it. Each handoff introduced delays, miscommunication, and quality issues. The company was turning away work because they couldn’t produce reports fast enough, and clients were frustrated by the turnaround time.

Solution

We built a multi-agent system where specialized AI agents collaborate to produce business reports. A research agent gathers data from web sources, databases, and internal documents. An analysis agent interprets the data, identifies patterns, and draws conclusions. A writing agent drafts the report in the company’s style. A review agent checks for accuracy, consistency, and completeness.

The agents work together in a structured workflow managed by LangGraph. The research agent’s output feeds the analysis agent, which feeds the writing agent, which feeds the review agent. If the review agent finds issues, it can send the report back to the writing or analysis agent for revision. A human analyst reviews the final output before it’s sent to the client, but their role shifts from creating the report to curating and refining the AI’s work.

Technologies

The system uses CrewAI for agent role definition and collaboration, LangGraph for workflow orchestration and state management, OpenAI for the reasoning behind each agent, and FastAPI for the service layer. Each agent has a specialized system prompt, a set of tools, and a defined output format.

The research agent has web search and database query tools. The analysis agent has data analysis and visualization tools. The writing agent has access to the company’s report templates and style guide. The review agent has a fact-checking tool and a quality rubric. LangGraph manages the state transitions between agents, handles revision loops, and ensures the workflow progresses toward a complete report. A FastAPI endpoint lets analysts trigger reports, monitor progress, and review outputs.

Results

Report generation became dramatically faster — what took 2-3 weeks now takes 1-2 days, with the AI agents doing the heavy lifting and the human analyst focusing on review and refinement. A large portion of manual research was eliminated, as the research agent gathers and organizes data that the analyst would have spent days collecting.

The company reduced analysis costs significantly because they could produce the same volume of reports with fewer people, and they stopped turning away work. Report quality actually improved, because the review agent catches inconsistencies and errors that human reviewers miss when they’re rushed. Client satisfaction increased, and the company was able to take on considerably more report projects without hiring additional analysts.

Screenshots

Multi-agent workflow with status of each agent

Generated business report with research citations

Architecture

The system is a LangGraph state machine with four agent nodes: research, analysis, writing, and review. Each node is a CrewAI agent with specialized tools and prompts. The state machine manages transitions: research → analysis → writing → review, with conditional edges for revision loops when the review agent identifies issues.

A FastAPI service exposes endpoints for triggering reports, monitoring progress, and retrieving outputs. The service manages a queue of report requests and runs multiple agent workflows concurrently. A shared memory store lets agents access each other’s intermediate outputs, so the writing agent can reference the research agent’s sources and the review agent can check the analysis agent’s conclusions. A human-in-the-loop interface lets analysts review, edit, and approve reports before delivery.