---> Agentic Design Patterns:

Building Blocks, Workflows, and Agents

Yeh text AI development process ko 3 levels mein todta hai:

1. Building Blocks (Bunyadi Hisse)

• Yeh chhoti chhoti cheezein hain jo AI ka base banati hain.
• Jaise "Augmented LLM" (jo retrieval, tools, aur memory use kar sakta hai).

2. Workflows (Kaam karne ka tareeqa)

• Yeh steps aur rules hain jo AI ko sahi tareeke se kaam karne mein madad dete hain.
• Jaise "AI jab ek tool use kare, toh pehle data retrieve kare".

3. Agents (Mukammal AI System)

• Yeh poora AI system hota hai jo independent tareeke se kaam kar sakta hai.
• Jaise ek customer support AI, jo messages read kare, relevant information dhoonde, aur sahi jawab de.

Nateeja (Conclusion)

Anthropic yeh keh raha hai ke AI ko directly ek advanced system banane ke bajaye, humein pehle chhoti chhoti cheezein banani chahiye (building blocks), phir workflows set karne chahiye, aur phir AI agents develop karne chahiye.

Yeh step-by-step approach AI development ko behtar aur asaan banati hai.

1. Augmented LLM (Building block):

Yeh text "Augmented LLM" ka concept aur uski importance explain karta hai. Chalo isay step by step samajhte hain:

1. Description (Tafseel)

• "Augmented LLM" ka matlab hai ek improved AI model, jo sirf apni built-in knowledge par depend nahi karta.
• Is model ko teen tareeqon se behtar banaya gaya hai:
  1. Retrieval – Yeh AI extra information dhoondh sakta hai (e.g., internet se ya database se).
  2. Tool Use – Yeh AI tools (e.g., calculator, translator, APIs) use kar sakta hai.
  3. Memory – Yeh AI pehle ki baatein yaad rakh sakta hai, taake ek jaise sawaalon ka behtar jawab de sake.
• Yeh features AI ko powerful banate hain, taake yeh apna kaam behtar kare, jaise:
  • LLM khud decide kar sakta hai ke kab search karna hai, konsi tool use karni hai, aur konsi baat yaad rakhni hai.
  • Aapko yeh features apni zaroorat ke mutabiq set karne chahiye taki system aapke kaam ke liye best perform kare.
  • In features ka interface simple aur achi tarah document hona chahiye, taki LLM inko asani se use kar sake.

Conclusion

Agar AI ko retrieval, tools aur memory ke saath integrate kiya jaye, toh yeh zyada smart, flexible aur useful ho jata hai. Lekin isko simple aur use case ke mutabiq design karna zaroori hai, taake yeh achay tareeke se kaam kare.

Augmented LLM Diagrame:

---

2. Workflow: Prompt Chaining:

Prompt chaining ek process hai jisme AI ek task ko chhote chhote steps mein tod kar solve karta hai. Har step ka output agle step ke input ke طور پر use hota hai.

Aap beech beech mein programmatic checks (gates) laga sakte hain taake process sahi track pe rahe.

Kab Istemaal Karein?

• Jab task ko chhoti, fix subtasks mein divide karna asaan ho.
• Yeh speed (latency) ko thoda kam kar ke accuracy improve karta hai. (yani time zayada lag sakhta hai kue ke kaam chote chote steps mien ho rha hai aur accuracy zyada hoti hai kue ke galtiya hone ke chance kaam hai)

Examples:

✅ Marketing copy likhna, phir usay kisi doosri zuban mein translate karna.
✅ Ek document ka outline likhna, phir check karna ke woh sahi hai, aur uske basis par poora document likhna.

Prompt Chaining Diagrame:

---

3. Workflow: Routing:

Routing ka matlab hai AI kisi input ko pehchankar usay sahi follow-up task ki taraf bhejta hai. Yeh AI ko specialized aur organized banata hai.

Agar yeh workflow na ho, toh ek hi model har input par kaam karega, jo performance ko girane ka sabab ban sakta hai.

Kab Istemaal Karein?

• Jab task ke alag-alag categories ho aur unko alag tareeke se handle karna behtar ho.
• Agar AI ya koi algorithm input ko sahi categorize kar sakta hai, toh yeh workflow zyada effective hota hai.

Examples:

✅ Customer service queries ko alag-alag process mein bhejna (e.g., general questions, refund requests, aur technical support).
✅ Asaan aur aam sawaal chhoti AI models (jaise Claude 3.5 Haiku) ko bhejna, aur mushkil aur unusual sawaal advanced models (jaise Claude 3.5 Sonnet) ko bhejna taake cost aur speed optimize ho sake.

Routing Diagrame:

---

4. Workflow: Parallelization:

Parallelization ka matlab hai AI ek hi kaam ko ek saath alag tareekon se kare taake tez aur behtar result mile.

Yeh do tareekon se hota hai:

1. Sectioning – Kaam ko chhoti chhoti parts (subtasks) mein tod kar ek saath solve karna.
2. Voting – Ek hi kaam ko baar baar karwana taake behtareen jawab mil sake.

Kab Istemaal Karein?

• Jab kaam ko tod kar tez solve karna ho.
• Jab multiple jawab check karke behtareen select karna ho.

Examples:

✅ Sectioning:

• Ek AI user ka question ka jawab de, doosra AI galaat ya inappropriate baat check kare.
• LLM ke performance ko judge karne ke liye har AI alag aspect check kare.

✅ Voting:

• Code checking: Multiple AI ek hi code ko check karein aur jo bari problem ho usay flag karein.
• Content checking: Multiple AI alag-alag nazar se check karein ke koi cheez inappropriate toh nahi.

Parallization Diagrame:

---

5. Workflow: Orchestrator-Workers

Orchestrator-Workers ka matlab hai ek AI (orchestrator) kaam ko chhoti subtasks mein todta hai aur multiple AI (workers) ko assign karta hai. Phir yeh AI sab results ko mila kar final answer deta hai.

Kab Istemaal Karein?

• Jab pehle se nahi pata hota ke kitne aur kis qisam ke chhote kaam karne padenge.
• Jab task complex ho aur har baar naya tareeka chahiye ho.

Example:

✅ Coding: AI multiple files ko alag-alag modify karein, kyunki har file ki zaroorat alag ho sakti hai.
✅ Search: AI alag-alag sources se information collect karein, analyze karein aur best result de.

Orchestrator Diagrame:

---

6. Workflow: Evaluator-Optimizer

Is workflow mein do LLM hote hain:

1. Optimizer (بہتر بنانے والا) – Pehla LLM kaam karta hai aur ek jawab generate karta hai.
2. Evaluator (جانچنے والا) – Doosra LLM us jawab ka review karta hai aur feedback deta hai.

Yeh process loop mein chalta hai jab tak acha result na mil jaye.

Misaal: Ek student ne essay likha (optimizer), aur teacher ne uska review kiya aur bataya ke kaha behtar ho sakta hai (evaluator). Student phir se improve karega jab tak essay best na ban jaye.

---

Kab istemaal karein?

✅ Jab humare paas clear evaluation criteria ho (matlab, hum jaan sakain ke acha result kaisa hoga).
✅ Jab feedback se result improve ho sakta ho.
✅ Jab AI khud feedback de kar result better bana sakay.

---

Example:

🔹 Literary Translation – AI translation kare, phir doosra AI check kare ke meaning aur nuances theek hain ya nahi.
🔹 Complex Search – AI search kare, phir doosra AI check kare ke jo mila wo kaafi hai ya aur search chahiye.

---

Summary:

Ye step-by-step behtari ka tareeqa hai (workflow hai), jisme ek AI kaam karta hai aur doosra usko behtar banata hai. Yeh tab tak chalta hai jab tak behtareen result na mil jaye.

Evaluator-Optimizer Diagrame:

-_______________________________________________________________________________________________________________________________________________-

-_______________________________________________________________________________________________________________________________________________-

Key Differences

Workflow	Kaam ko kaise divide karta hai?	Fixed ya Flexible?	Kab istemaal karein?
Prompt Chaining	Pehla AI kaam kare, doosra uska output use kar ke agla step kare.	Fixed	Jab step-by-step AI tasks ki zaroorat ho.
Routing	Pehle se fixed rules ke mutabiq alag models ko kaam deta hai.	Fixed	Jab alag-alag qisam ke kaam alag experts (LLM) ko dene ho.
Parallelization	Kaam ke multiple parts ek saath process hotay hain.	Fixed	Jab speed aur accuracy improve karni ho.
Orchestrator-Workers	Ek AI dynamically decide karta hai ke kaam kaise todna hai.	Flexible	Jab har dafa naye rules ke mutabiq kaam todna ho.
Evaluator-Optimizer	Pehla AI jawab banata hai, doosra AI feedback de kar improve karta hai.	Flexible	Jab iterative improvement chahiye aur AI feedback se behtar result de sakta ho.

---

Final Summary

• Prompt Chaining – Jab AI ko ek structured sequence mein kaam karwana ho.
• Routing – Jab kaam predefined categories mein divide ho sakta ho.
• Parallelization – Jab jaldi aur zyada kaam saath-saath karna ho.
• Orchestrator-Workers – Jab task unpredictable ho aur naye rules ke mutabiq divide karna ho.
• Evaluator-Optimizer – Jab quality improve karni ho feedback ke zariye.