The Dial, aur paanch Patterns

Lever 01 ko paas se: har point par decide karna ki aapka code drive karta hai ya model — aur workflow side jin paanch shapes mein aata hai unhe name dena.

Lesson 1 ne aapko map diya aur uska master lever: control flow — agla step kaun decide karta hai, aapka code (ek workflow) ya model (ek agent). Apne intuition mein karne waali sabse zaroori correction yeh hai: yeh choice ek baar poore system ke liye nahi hoti. Yeh ek dial hai jo aap har decision point par set karte ho.

Ek production “agent” aam taur par ek mostly-deterministic program hota hai jisme model kuch high-leverage spots par dial-in kiya hota hai, aur jahaan reliability code se sasti milti hai wahaan dial out kiya hota hai.2 Is lesson mein aapka kaam: dial ko deliberately set karna seekhna, aur workflow end jin paanch named shapes mein aata hai unhe pehchaanna.1

Wo default jo cleverness ko harata hai

Anthropic ki headline finding, dozens production deployments survey karne ke baad, lagbhag anticlimactic hai:

Consistently, the most successful implementations weren't using complex frameworks… they were building with simple, composable patterns. Find the simplest solution possible, and only increase complexity when needed… add complexity only when it demonstrably improves outcomes. — Anthropic, Building Effective Agents

Toh dial ki ek default position hai, aur default hai “code ki taraf.” Model end ki taraf har click aap tab earn karte ho jab dikhaate ho ki wo aisa kuch deta hai jo ek deterministic path nahi de sakta tha. Kisi autonomous agent tak haath badhaane se pehle, do sawaal poochho:

Paanch workflow patterns

Dial ke workflow side par jo bhi hai wo in paanch patterns ka composition hai. Inhe ek vocabulary ki tarah seekho: jab koi problem aaye, toh aap shape ko name dena chaahte ho, usse invent nahi karna.1

Inme se do, code mein

Routing aur evaluator–optimizer wo do hain jinke liye aap sabse zyada haath badhaoge. Dono structure ko code mein rakhte hain aur model ko leash par — bilkul wahi discipline jise Lesson 1 ne kaha tha “own your control flow”:

route = model(classify_prompt(query)).label    # model sirf ek label return karta hai
handler = {                                     # switch aapka hai, model ka nahi
    "refund":    handle_refund,
    "technical": handle_technical,
    "general":   handle_general,
}[route]
return handler(query)                           # har category ke liye specialised path

draft = model(generate_prompt(task))
for _ in range(MAX_ROUNDS):                     # bounded loop — dekho Lesson 4
    verdict = model(critique_prompt(draft))     # ek doosra call draft ko grade karta hai
    if verdict.passes:
        return draft
    draft = model(revise_prompt(task, draft, verdict.feedback))
return draft                                    # budget ke andar best effort

Agents apni keemat kahaan vasoolte hain

Aap poora dial agent tak tab le jaate ho jab “open-ended problems where it’s difficult or impossible to predict the required number of steps.”1 Wo flexibility cost, latency, aur galtiyon ke liye ek zyada bade blast radius se khareedi jaati hai — aur isiliye baad ke levers (context, budgets, error recovery) hain: wahi guard-rails hain jo dial ke agent end ko use karne layak safe banaate hain.

Recall check

Retrieve karo, dobara mat padho. Har scenario ko memory se uske pattern par map karo.

Sources

1. Anthropic — Building Effective Agents. Workflow vs. agent; the five patterns and when to use each; “add complexity only when it demonstrably improves outcomes.”
2. HumanLayer — 12-Factor Agents. Factor 8: own your control flow — most of a reliable agent is ordinary code with the model at key points.