Multi-Agent

A modular LangGraph team that answers a task by routing work through specialist agents.

The supervisor decides who works next. A researcher gathers notes, a writer drafts the answer, and a reviewer approves it or sends it back for revision.

START → supervisor → researcher → writer → reviewer → finish → END
                         ↑                      |
                         └──── revise (optional)┘

Requirements

• Python 3.10+
• A Metis or OpenAI-compatible API key

Setup

pip install -r requirements.txt

Copy the example env file (or edit the existing .env) and put a real API key in it:

copy .env.example .env

METIS_API_KEY=
OPENAI_API_KEY=
OPENAI_BASE_URL=https://api.metisai.ir/openai/v1
LLM_MODEL=gpt-4o-mini
LLM_TEMPERATURE=0
MAX_REVISIONS=2
RECURSION_LIMIT=25

Set METIS_API_KEY or OPENAI_API_KEY. A placeholder such as your-metis-api-key will not work.

The default API base URL is https://api.metisai.ir/openai/v1. Override it with OPENAI_BASE_URL or METIS_BASE_URL if you use another endpoint.

Usage

One-shot:

python main.py "Explain how LangGraph supervisor agents work"

Interactive session:

python main.py

Type a task, or exit / quit to leave.

Hide node progress logs:

python main.py --quiet "Summarize the difference between agents and workflows"

How it works

1. Supervisor reads shared state and picks the next worker: researcher, writer, reviewer, or finish.
2. Researcher collects structured notes for the task.
3. Writer turns those notes (and any review feedback) into a draft.
4. Reviewer approves the draft or requests changes.
5. Finish delivers final_answer.

If structured routing fails, the supervisor falls back to simple rules (research first, then write, then review). Revisions stop after MAX_REVISIONS (default 2) so the graph cannot loop forever.

Project layout

config/          Env-based settings
models/          Shared AgentState and Pydantic schemas
prompts/         System and user prompts
agents/          One module per worker
graph/           Graph wiring, routing, and run_task
utils/           LLM factory and logging
main.py          CLI entry point

Call the team from code:

from graph.runner import run_task

result = run_task("What is a LangGraph state graph?")
print(result["final_answer"])

Configuration

Variable	Default	Purpose
METIS_API_KEY	—	Preferred API key
OPENAI_API_KEY	—	Fallback API key
OPENAI_BASE_URL / METIS_BASE_URL	https://api.metisai.ir/openai/v1	OpenAI-compatible endpoint
LLM_MODEL	gpt-4o-mini	Chat model
LLM_TEMPERATURE	0	Sampling temperature
MAX_REVISIONS	2	Reviewer rewrite limit
RECURSION_LIMIT	25	LangGraph step limit

Extending

1. Add a node in agents/.
2. Register it in graph/builder.py.
3. Teach the supervisor the new name in models/schemas.py and prompts/templates.py.
4. Update graph/routing.py so the supervisor can send work there.