Chapter 43

AI এজেন্ট

AI Agents

🎬 LLM থেকে Autonomous System

AI Agent = LLM + tools + memory + planning loop। শুধু একটি উত্তর নয় — agent নিজেই সিদ্ধান্ত নেয়, tool কল করে, ফলাফল observe করে, পরের পদক্ষেপ ঠিক করে। লক্ষ্য পূরণ পর্যন্ত চলে।

Agent Loop (ReAct)

while not done:
    thought  = llm("আমি এখন কী করব? কারণ?")
    action   = parse_tool_call(thought)
    if action == FINISH: break
    obs      = run_tool(action)
    memory.append((thought, action, obs))
return final_answer

একটি Agent-এর উপাদান

LLM brain — reasoning ও decision।
Tools / Functions — search, calculator, code-exec, DB, API।
Memory — short-term (context), long-term (vector store)।
Planner — goal decompose, sub-task।
Controller / Loop — orchestration ও stopping rule।

Tool / Function Calling

from openai import OpenAI
client = OpenAI()

tools = [{
  "type": "function",
  "function": {
    "name": "get_weather",
    "description": "Get current weather for a city.",
    "parameters": {
      "type": "object",
      "properties": {"city":{"type":"string"}},
      "required": ["city"]
    }}}]

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role":"user","content":"Dhaka-এর weather?"}],
    tools=tools, tool_choice="auto")

call = resp.choices[0].message.tool_calls[0]
result = get_weather(**json.loads(call.function.arguments))

Agent Frameworks

Framework

শক্তি

LangGraph

state-machine, production-ready

LlamaIndex Agents

RAG-centric agent

CrewAI

role-based multi-agent

AutoGen (MS)

conversational multi-agent

OpenAI Agents SDK

simple, tool-native

LangGraph উদাহরণ

from langgraph.prebuilt import create_react_agent
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool

@tool
def calc(expr: str) -> str:
    "Evaluate a math expression."
    return str(eval(expr))

agent = create_react_agent(ChatOpenAI(model="gpt-4o-mini"), tools=[calc])
out = agent.invoke({"messages":[("user","12*13 + 7^2 কত?")]})
print(out["messages"][-1].content)