ReAct Agent Flow

A local multi-agent system demonstrating the ReAct (Reasoning + Acting) loop using LangGraph, FastAPI, and Ollama.

Architecture

The supervisor agent reasons over tool results iteratively — calling specialist agents, reading their outputs, and deciding whether to delegate further or return a final answer. Max 3 iterations per query.

User query
  → Supervisor (iteration 1) → research / coding / summarizer / critic tools
  → Supervisor (iteration 2) → reads all results → critic if needed → final answer

Stack

• Backend: FastAPI + LangGraph + LangChain
• LLM: Ollama (llama3.1:8b by default, configurable via .env)
• Frontend: Streamlit (live WebSocket event stream)
• Containerization: Docker Compose

Run

docker compose up --build

Open http://localhost:8501. On first run, Ollama pulls the configured model automatically (~4.7 GB).

Configuration

Copy .env.example and set your model:

MODEL=llama3.1:8b
OLLAMA_BASE_URL=http://ollama:11434

Agents

Agent	Role
research	Factual questions, explanations, comparisons
coding	Writing, debugging, explaining code
summarizer	Merging multiple outputs into one answer
critic	Reviewing and improving a draft

The supervisor decides which agents to call and in what order using native LLM tool calling (bind_tools), not text parsing.