Deploying Lár Agents (Production Guide)

Deploying Lár Agents (Production Guide)

One of the most common questions is: "How do I deploy this?"

Lár is an Engine (like PyTorch or Flask), not an Application (like ChatGPT). To deploy it, you simply wrap it in a standard Python web server.

1. The Strategy: "Headless" Engine

You should run your agent as a stateless REST API.

• Input: JSON (The User Task)
• Output: JSON (The Final State)
• Audit Log: Save the result_log to a database (Postgres/Mongo) for compliance.

2. Using FastAPI (Recommended)

Deep within the examples/ folder is Example 4. It is a complete, copy-pasteable implementation of a Lár Agent running on FastAPI.

examples/basic/4_fastapi_server.py

Quick Start

1. Install FastAPI:

   pip install fastapi uvicorn
   

2. Paste this code:

   import uvicorn
   from fastapi import FastAPI
   from lar import LLMNode, GraphExecutor
   
   app = FastAPI()
   executor = GraphExecutor()
   
   # Define a simple agent
   agent = LLMNode(
       model_name="gpt-4o", 
       prompt_template="Echo: {task}", 
       output_key="response"
   )
   
   @app.post("/run")
   def run_agent(task: str):
       # Run standard Lár execution
       result = list(executor.run_step_by_step(agent, {"task": task}))
       return result[-1]["state_snapshot"]
       
   if __name__ == "__main__":
       uvicorn.run(app, port=8000)
   

3. Deploy: Run this script on Heroku, AWS Lambda, Railway, or Render just like any other Python app.

3. Why not LangServe?

Frameworks like LangChain force you to use their proprietary "Serving" layers (LangServe) which often lock you into their ecosystem.

By using standard FastAPI, you essentially "own" the deployment. You can:

• Add custom Authentication (OAuth2, JWT).
• Rate Limit users by IP.
• Save logs to your own SQL database.
• Integrate with existing Stripe payment flows.