Class 12 · 12.4

Agents aur sub-agents (aur Codex ki jhalak)

Ab haftay ka markazi mauzu: agents aur sub-agents. Hum “multi-agent” ke hype ko asal sona (sub-agents) se alag karenge, apna reviewer banayenge, kaam poori tarah doosre AI ko denge aur aane wali agent-teams ki duniya ka preview dekhenge.

🎯 Maqsad: doosre agents ko kaam dena⏱ Parhein: 16 min🤖 Qisam: hands-on

“Multi-agent” kuch pur-asrar lagta hai. Is ka ek hissa waqai mamooli hai; qeemti hissa — sub-agents — asal taqat yahin hai. Aayiye tartib se dekhte hain.

01“Multi-agents”: saada haqeeqat

Multi-agent ka sab se saada matlab hai… kai terminals khol kar kai Claudes chalana. Ek frontend banaye, ek backend aur ek tests. Agar boundaries saaf hon to yeh parallel kaam kar sakte hain.

Ehtiyati demo

Patli si plan.md ke saath teen independent agents chalana aksar fiasco hoga — shared bunyaad kam, agents ek doosre se be-khabar aur pieces aapas mein fit nahi. Is se pata chalta hai ke “multi-agent” ka kya matlab hai aur parallel karne se pehle coordination (12.2 ki shared files) kyun zaroori hai.

02Bilkul mukhtalif agent: Codex CLI

Aap ka doosra agent Claude hona zaroori nahi. Baat ko wazeh karne ke liye Codex laate hain — OpenAI ka Claude Code jaisa tool, jo command line se chalta hai. Install karein aur phir codex exec se baghair interaction chalayein:

terminal

npm install -g @openai/codex     # or: brew install codex
codex                                # interactive; first run does an OAuth login

# the magic: run a one-shot prompt, no chat
codex exec "review planning/plan.md and write feedback to planning/review.md"

Yeh badi baat kyun hai

codex exec poore doosre AI ko ek shell command ki tarah chalata hai. Is ka matlab hai ke koi bhi agent script se launch ho sakta hai — aur kuch der mein Claude Code ke andar se bhi. Is se waqai mixed multi-agent workflows ka darwaza khulta hai. (Codex zaroori nahi — doosra claude bhi itna hi kaam karta hai; different-LLM wala twist bas mazedaar hai.)

03Sub-agents: asal sona

Sub-agent alag agent hota hai jo ek task leta hai, apne alag context mein chalata hai aur main Claude Code ko sirf result wapas deta hai. Aap built-in agents (Explore, Plan) ko mukhtalif rang mein dekh chuke hain.

Shor bhari aana-jaana sub-agent ke context mein hoti hai. Main agent ko sirf saaf jawab nazar aata hai — poora nuqta yahi hai.

Isay kyun use karein? Chaar wajahen: context bachana (sab se badi baat — main window ko shor se bachana), parallelism (ek waqt mein kai), focus (ek kaam ke liye perfect prompt), aur cost/speed (saade tasks ke liye Haiku jaisa sasta model).

04Reviewer sub-agent banayein

Sub-agents .claude/agents/ mein thore front matter wali markdown files ki shakal mein hote hain. Yeh reviewer hai — aur mazedaar version asal review Codex ko deta hai:

.claude/agents/reviewer.md

---
name: reviewer
description: Carry out a comprehensive review when requested.
# tools: and model: are optional — omit to inherit the caller’s
---
You carry out a comprehensive review of plan.md when requested,
using a different AI. You MUST run the following shell command and
do NOT review it yourself:

codex exec "review planning/plan.md, write feedback to planning/review.md"

Slash se call nahi karte

Command ke baraks aap /reviewer type nahi karte. Saadi zaban mein is ke liye kehte hain — behtar jumla is ka naam saaf leta hai: “planning/plan.md review karne ke liye reviewer sub-agent use karein.” Claude isay launch karne ka faisla karta hai aur poora review alag chalta hai.

claude

> use the reviewer sub-agent to review planning/plan.md

# main Claude spawns the sub-agent → it runs `codex exec` →
# Codex writes planning/review.md → only the result returns.
# /context shows the main window is still almost empty.

Isay aam banayein

“review plan.md” ki jagah “last commit ke baad ki tamam tabdeeliyon ka review karein” likhein aur aap ke paas reusable change-reviewer hai — real multi-agent orchestration ka beej, aur bilkul woh cheez jo aglay qadam mein hook se automatically chalayenge.

05Sub-agents aur agent teams

Chhota sa preview taake baad mein terms aap ko uljhayen nahi (agent teams wild day par aayengi):

	Sub-agents	Agent teams (experimental)
Shape	One main agent delegates one task, gets the result back.	A whole group of agents running collaboratively at once.
Communication	Only through the main agent.	Agents can talk to each other directly.
Lifespan	Short: take the task, do it, return.	Long-running presence in your “team”.
Example	A reviewer that reads a file and reports back.	A tester that gives feedback straight to the frontend and backend agents.

06Sub-agents ka jaiza

Khoobiyan

Inhein kyun use karein

Main context bachayein, kaam parallel chalayein, self-correct karein (review/test agents), aur ek perfect prompt ko ek kaam par focus karein.

Kharch

Badlay mein kya dena hai

Zyada moving parts (ghalati ke zyada imkaan), boundary/interface ki ghalatiyan, ek agent bhatke to barhti hui errors, aur kabhi kabhi extra churn ki wajah se zyada kharch.

Hamesha saste nahi

Sub-agents ko saste models se paisa bachane ka tareeqa kaha jata hai — lekin self-correcting review loops ka matlab hai ke kul calls zyada hoti hain. Behtar results, haan; guaranteed bachat, nahi. Jahan context ya parallelism ka asal faida ho wahin use karein.

✓ Ahm baatein

“Multi-agent” ka matlab kai terminals bhi ho sakta hai — coordination count se zyada aham hai.
codex exec doosre AI ko ek shell command ke taur par chalata hai — scripts aur Claude dono se launch ho sakta hai.
Sub-agent isolated context mein task chalata hai aur sirf result wapas deta hai — aap ki window bachti hai.
Inhein .claude/agents/*.md mein define karein; “use the … sub-agent” keh kar bulayein.
Agent teams (baad mein) mil kar baat karne wale kai agents chalati hain; sub-agents simple aur task-focused hote hain.