Lesson 63 — Beginner

What is an AI Agent?

Agent ek aisa LLM system hai jo khud decide karta hai ki agla step kya hoga — tum sirf goal dete ho, raasta model chunta hai. Is lesson mein agent loop, call vs chain vs workflow vs agent ka farak, aur wo decision rule seekhoge jisse pata chalega ki problem ko agent chahiye ya ek simple workflow kaafi hai.

Agent loop — goal se finish tak

Loop ka poora cycle ek line mein: goal → decide → act → observe → repeat → finish. Har iteration ek model call hai, har tool call ek action, aur har tool result ek observation jo agle decision ko shape deta hai.

def run_agent(goal, tools, llm):
    messages = [{"role": "user", "content": goal}]
    while True:  # loop chalta rahega jab tak finish na aaye
        reply = llm(messages, tools=tools)
        if not reply.tool_calls:  # action nahi chahiye = jawab final
            return reply.content
        messages.append(reply)    # decide kiya — ab act
        for call in reply.tool_calls:
            result = run_tool(call.name, call.args)
            messages.append(tool_msg(call.id, result))  # observe

Call vs chain vs workflow vs agent

shaped          | agla step kaun tay karta hai
----------------+-------------------------------------------
call            | koi nahi — ek prompt, ek jawab
chain           | code jo tumne likha (A -> B -> C fixed)
workflow        | code-based branches/rules — path bhi tumhara
agent           | model khud — tools aur observations dekh kar

Ek hi spectrum hai: jitna zyada control tumhare code mein, utna kam model ke paas. Call sabse seedha hai, agent sabse aazaad. Beech ke forms (workflow + ek agent step) hi production mein sabse zyada chalte hain.

• Call: classification, extraction — ek sawaal, ek jawab, koi loop nahi.
• Chain: prompt → parse → prompt — order tumne likha, model sirf steps bhar raha hai.
• Workflow: fixed DAG with conditionals — jaise "agar invoice high hai toh approval branch, warna auto-approve".
• Agent: koi fixed path nahi — decision model ka hai, tum sirf tools aur rules dete ho.

Agent ke components

agent = {
  model:        kaun sochta hai (LLM choice)
  instructions: kaise sochna hai (system prompt, rules, tone)
  tools:        kya kar sakta hai (search, db, api, sandbox code)
  state:        kya yaad rakhta hai (history, plan, memory)
  loop:         decide -> act -> observe -> repeat
  stopping:     kab rukna hai (final answer, max steps, budget)
}

• Model + instructions — brain aur niyam. Instructions mein hi bata do ki kab tool use karna hai aur kab seedha jawab dena hai.
• Tools — duniya ke saath contact. Tools bekaar toh agent bekaar.
• State/memory — current conversation window, task plan, purane facts.
• Control loop + stopping conditions — loop ka code, max steps, budget, finish rule. Bina rukne ke niyam ke agent infinite chal jaata hai.

Kab workflow hi behtar hai

Decision rule: agar agla step tum apne code mein likh sakte ho, toh likh do. Deterministic workflow simpler, sasta aur zyada reliable hota hai — har decide-step ek extra LLM call hai jo latency aur token dono khaata hai.

• Workflow better: fixed extraction pipelines, standard RAG chains, batch jobs, forms validation — path known hai, agent ki zarurat nahi.
• Agent useful: research jahan source ke results decide karte hain agla query, support jahan har ticket ki investigation alag ho, coding tasks jahan test feedback loop chalta hai.
• Hybrid: workflow ke andar ek agent step — jaise "draft reply likho" wala step. Control bhi milega, flexibility bhi.
• Cost check: agent ka har turn ek model call — agar 3 fixed calls kaafi hain toh 10-turn loop kyun chalana.

Exercise

Common mistakes

• Har problem pe agent banana: Jab path fixed hai toh workflow simpler, sasta aur reliable hai. Pehle workflow banao, agent sirf tab jab decide karna genuinely model ka kaam ho.
• Agent ko bina stopping rule ke chhodna: Max steps, token budget aur explicit finish ke bina loop ghoomta rehta hai. Har agent ke saath stopping conditions pehle se define karo.
• Agent = chatbot samajhna: Chatbot aksar ek call hai — agent ek LOOP hai jisme model tools se baar baar interact karta hai. Farak "agla step kaun decide karta hai" mein hai.

workflow: order -> fixed menu -> bill
agent:    order -> waiter sochta hai -> poochhta hai
          -> nayi jaankari -> rasta badla -> repeat -> serve