Subagent ek alag context hai jiske apne instructions, apne tools aur aksar apna model hota hai — wo useful hi isliye hai ki jo kuchh wo padhta hai, wo main conversation mein nahi ghusta. Ye alagav kis had tak saarthak hai, ye do fields tay karte hain. Description kaam sounpe jaane ki wajah banta hai aur aapke tay kiye baaki sabhi agents se hod karta hai, isliye dhundhla description ya to kabhi nahi chalta ya lagataar chalta hai. Tools list kisi seemit agent ko seemit banaye rakhne ka sabse sasta tareeka hai: jo sirf padh aur khoj sakta hai, wo bhatakkar edit nahi karne lagega. Ye builder file likhta hai aur dono jaanchta hai — us jodi samet jo adhiktar logon se chhoot jaati hai: likhne ka adhikaar, padhne ka nahi.
Features
Description mein delegation karane wale phrase ki jaanch
Tools yaad se type karne ke bajaye asli list se chune jaate hain
Namespace wale MCP tools bhi built-in tools ke saath accept
Agent file likh sakta ho par padh na sake to warning
Model ka chunav "kab saarthak hai" ke hisaab se samjhaya gaya
Project ya personal scope, saath mein upyukt path
Inherit kiya gaya model file mein likha nahi jaata
Seedhe use layak .md file ke roop mein download
Subagent Builder kaise use karein
Agent ka naam rakhiye aur likhiye ki kaam use kab sounpa jaye
Sirf wahi tools chuniye jinki use zaroorat hai
Model chuniye, ya session wala hi inherit karne dijiye
System prompt likhiye aur file download kijiye
Example
Input
tools: [Read, Grep, Glob]
Output
description: Checks a database migration for problems that only show up in production. Use proactively whenever a migration file is added.
"Use proactively after…" wahi hai jo agent ko bina kahe chala deta hai.
Common errors aur troubleshooting
Agent ko kabhi kuchh delegate hi nahi hota. — Description ko "kab" kehna hoga, sirf "kya" nahi. "Use when…" ya "Use proactively after…" jod dena aam taur par kaafi hai.
Do agents aapas mein gaddmadd hote rehte hain. — Unke descriptions ek-doosre par chadh rahe hain. Har description alag topic nahi, alag sthiti named kare.
Kisi tool ka naam anjaan bataya ja raha hai. — Naam case-sensitive hain aur capital se shuru hote hain: Read, read nahi. MCP tools mcp__server__tool roop mein hote hain aur waise hi accept hote hain.
Agent wo cheezein badal deta hai jinki ummeed nahi thi. — Tools list aur sankri kijiye. Jis agent ke paas Write aur Edit nahi, wo prompt chahe jo kahe, kuchh badal nahi sakta.
Aksar pooche jaane wale sawaal
Seedhe kaam karne ke bajaye subagent kab behtar hai?
Jab kaam "bahut padho, thoda nikalo" wala ho: codebase khangalna, dheron files jaanchna, khojbeen. Jo kuchh wo padhta hai, wo usi ke context mein reh jaata hai. Main session se sankre tools par chalane ka tareeka bhi yahi hai.
Subagent ke liye kaun-sa model?
Inherit karna achha default hai. Jahaan faisle se zyada jaanch hai, aise yantrik kaamon ke liye tez aur sasta model saarthak hai; jahaan vivek chahiye, wahaan zyada saksham. Tulna har vikalp ke saath likhi hai.
Kya tools hamesha seemit karne chahiye?
Seemit karna sasta aur palatne yogya hai, aur kisi focused agent ko focused rakhne ka sabse bharosemand tareeka — prompt mein kehne se kahin zyada. List khaali chhodne par sab kuchh inherit hota hai, jo saamanya kaamon wale agent ke liye theek hai.
Project scope ya personal?
Project scope agent ko repository ke saath rakhta hai, to poori team ko ek hi milta hai. Personal aapke saath projects mein chalta hai. Path chunav ke saath badalta rehta hai.
Ye skill se kaise alag hai?
Skill maujuda context mein instructions jodti hai. Subagent kahin aur chalta hai, apne context aur apne tools ke saath, aur lautkar batata hai. Koi kaam kaise ho ye badalna hai to skill; us kaam ko apni conversation se bahar rakhna hai to subagent.