Multi-Agent Competitive Intelligence
A Supervisor-Worker multi-agent system built in LangGraph that accepts a public company name and produces a structured six-section competitive intelligence brief using parallel specialist agents and live web search.
- Context
- MSBA — Leveraging LLM Productivity, UT Austin
- Stack
- Python, LangGraph, NVIDIA NIM / OpenAI, DuckDuckGo Search
Architecture
Supervisor-Worker architecture — Supervisor delegates to specialist Worker agents, each responsible for one section of the brief
How It Works
The system uses a Supervisor agent that orchestrates a set of Worker agents, each responsible for a distinct section of the brief: company overview, competitive landscape, financial health, recent news, SWOT analysis, and strategic outlook. Workers run in parallel where possible and report back to the Supervisor for synthesis.
If live external services are unavailable, the workflow returns a partial-but-honest brief with disclaimers rather than failing silently — a deliberate design choice for production reliability.
Key Design Decisions
Supervisor-Worker Architecture
A central Supervisor node manages task delegation, tracks worker state, and assembles the final output — keeping worker agents focused and independently testable.
Provider-Agnostic LLM Layer
The system supports multiple LLM providers (NVIDIA NIM, OpenAI) via environment config, making it easy to swap models without changing orchestration logic.
Graceful Degradation
Workers that fail due to rate limits or search outages return honest partial results rather than propagating errors, preserving overall brief quality.
Structured Output
The final brief is rendered as both Markdown and PDF, making outputs immediately shareable without post-processing.