Skip to main content

Personal Agent Harnesses

Open-Source AI Employees Jinhein Aap Khud Run Aur Own Karte Hain

Pichhle section ne aapko general agents diye — Claude Code, OpenCode, Cowork, aur OpenWork — aur unhein drive, direct, aur loop karna sikhaya. Powerful, lekin mehdood. Aik general agent us session ke andar chalta hai jo aap khud kholte hain. Woh tab tak kaam karta hai jab tak aap dekh rahe hain, aur jab session khatam hota hai, us ka zyada tar working context bhi us ke saath khatam ho jata hai. Runtime aap ho. Laptop band karein aur worker ka wujood khatam.

Yeh section us ceiling ko hata deta hai.

Aik agent harness woh software hai jo aik model ko aisa worker bana deti hai jo aap ke baghair chalta hai. Is saal ki harness-engineering literature ne isay seedha keh diya: aik agent yaani model jama harness. Model sawalon ke jawab deta hai. Harness woh cheez hai jo usay continuously chalne, har session ke darmiyan jo seekha use yaad rakhne, aur action lene ke liye tools call karne deti hai. Yeh woh layer hai jo har us single model se zyada zinda rahegi jise aap is mein plug karte hain, aur theek isi liye platform vendors ab is par lar rahe hain.

Yahi woh line hai jis ke us paar yeh section aapko le jata hai: aik aise agent se jise aap drive karte hain, aik aise agent tak jise aap own karte hain.

Har harness kis cheez se bani hoti hai​

OpenClaw aur Hermes is aik baat ke siwa qareeban har cheez par ikhtilaf rakhte hain. Dono mein se kisi aik ko bhi tor kar dekhein to wahi anatomy nazar aati hai:

  • Runtime: agent ko tasks ke darmiyan zinda rakhti hai, sirf chat ke douran nahin.
  • Gateway: messages andar aur bahar le jaati hai, taake agent aap tak wahan pohnch sake jahan aap pehle se hain.
  • Memory: sessions ke aar paar mehfooz rehti hai, taake agent har din kal se zyada jaanta hua start kare.
  • Tools: woh external capabilities jinhein agent cheezein actually karne ke liye call karta hai (MCP servers, APIs).
  • Skills: portable expertise jo agent uthata aur dobara use karta hai (open agentskills.io format, cross-runtime).
  • Identity: agent kis ke taur par chalta hai, aur woh identity kin cheezon ko chhoo sakti hai.
  • Policy & observability: woh kya kar sakta hai, aur jo us ne kiya us ka record.

Is shape ko aik dafa seekh lein aur dono crash courses isi aik frame se latak jate hain. Aage aap jo bhi add karte hain (aik naya channel, aik naya skill, aik scheduled job) woh in saat mein se hi kisi aik ka aur hissa hai. In mein se do — skills aur connectors (oopar diye gaye tools) — portable hisse hain: open formats par bane hone ki wajah se yeh claude.ai, un general agents, jinhein aap ne drive kiya, aur us harness ke aar paar chalte hain jo aap yahan banate hain, is liye aap aik worker ko jo sikhate hain woh kisi aik surface tak mehdood nahin rehta.

Aik hi layer par do daanv​

OpenClaw aur Hermes do tools nahin jo aik hi kaam kar rahe hon. Yeh do daanv hain is baat par ke harness ka kaun sa hissa control point hai. Farq emphasis ka hai, exclusivity ka nahin. OpenClaw ke paas memory aur skills hain; Hermes beeson channels par baat karta hai. Lekin center of gravity hi hai jo har aik ko define karta hai.

Aik fork diagram. Oopar aik single box, "Wahi harness," saat shared hisse listate hai: runtime, gateway, memory, tools, skills, identity, policy. Yeh do boxon mein bant jata hai. Baayan, "OpenClaw — gateway par daanv," pehle breadth: aik agent kai channels par jaise WhatsApp, Telegram, Discord, aur Slack. Daayan, "Hermes — memory par daanv," pehle depth: yeh hafton mein aapko seekhta hai aur us infrastructure par chalta hai jo aap ki apni hai. Aik caption: wahi saat-hisson wali layer par mukhtalif control point, aur kuch bhi aapko dono chalane se nahin rokta.

OpenClaw gateway par daanv lagata hai. Pehle breadth. Aik agent WhatsApp, Telegram, Discord, Slack, aur is se aage tak aik hi jagah se jawab deta hai, jise aik bara community skills marketplace support karta hai. Yeh woh project hai jis ne sabit kiya ke personal AI Employees real hain aur log unhein chahte hain. Is ki taqat reach hai.

Hermes memory par daanv lagata hai. Pehle depth. Aik agent aap ka codebase, aap ke conventions, aur aap ke maazi ke decisions hafton tak rakhta hai, kisi mushkil task ke baad naye skills develop karta hai, aur kaam karte hue unhein refine karta hai. Yeh aise infrastructure par persistently chalne ke liye bana hai jo aap ki apni hai. Is ki taqat yeh hai ke yeh aap ko seekhta hai.

Wahi tradeoff jo aap cloud se pehle se jaante hain: aik broad, convenient gateway bamuqabla aik deep, self-managed worker jo waqt ke saath barhta hai. Aap is section se is qabil ho kar niklenge ke default ke bajaye soch samajh kar choose karein, aur dono ko run kar sakein, kyun ke yahan kuch bhi aapko either/or par majboor nahin karta.

Ownership hi asal baat kyun hai​

Dono courses ke neeche aik khamosh daleel hai, aur yahi wajah hai ke yeh section General Agents ke andar aik footnote ke bajaye us ke peer ki tarah mojood hai.

Jab aik harness aap ke agent ki memory, identity, aur runtime ko zinda rakhti hai, to sawal "is maheene kaun sa model sab se smart hai" hona band ho jata hai aur ban jata hai "jo worker mein train kar raha tha us ka maalik kaun hai." Platform vendors yeh jaante hain. Microsoft ka Scout aur Nvidia ka NemoClaw inhi harnesses ko governance aur identity mein wrap karte hain: convenient, production-ready, aur vendor ki milkiyat. Woh agent jis ne aap ki aadaton ka aik saal seekh liya hai, sab se bara switching cost hai. Memory, channel reach se zyada, lock-in ki woh paaedaar shakal hai jis par platform vendors daanv laga rahe hain.

Yeh section doosra raasta sikhata hai: woh open-source harness jise aap khud run karte hain, jahan runtime, memory, aur identity aap ki hain. Yahan ownership lafzi hai, cost samet: jo harness aap khud run karte hain woh model ko aap ki apni API access ke zariye call karta hai — metered, har call par paid — na ke kisi bundled claude.ai subscription par sawaar ho kar, is liye runtime, memory, identity, aur bill sab aap ke hain. Yahi woh trade hai — full control, full responsibility. Is liye nahin ke vendor wraps ghalat hain (kisi regulated enterprise ke liye woh aksar theek hote hain) balki is liye ke aap us tradeoff ka faisla tab tak nahin kar sakte jab tak aap ne self-owned version ko apne haathon se bana aur thaam na liya ho.

Yeh section shared path par aik step hai, side trip nahin. Yeh aik agent use karne (General Agents section) aur apne mode par commit karne (Mode 1 ya Mode 2) ke darmiyan trunk mein baitha hai: specialize karne se pehle har koi yahan khud ko apna persistent agent deta hai. Aap install ko coding agent (Claude Code ya OpenCode) ke zariye drive karte hain, jo setup aap ke liye kar deta hai, is liye yeh step aap ke liye open hai chahe aap khud code likhte hon ya nahin. Aur jo harness aap yahan banate hain, wahi cheez Mode 1 -> Mode 2 handoff baad mein apni do roads mein se aik ke taur par name karta hai, is liye us fork tak pohanchne se pehle aap us cheez ko already own kar chuke honge.

Yahan se shuru karein​

Dono crash courses aik prerequisite share karte hain: aap harness ko aik general agent ke zariye drive karte hain. Woh Claude Code ya OpenCode jo aap ne Agentic Coding mein seekha, wahi cheez hai jo harness ko aap ke liye install, configure, aur operate karti hai. Pichhle section ka general agent is section mein worker ka installer ban jata hai. Agar aap ne abhi tak nahin ki to pehle General Agents section karein.

  • OpenClaw with General Agents: 90 minute, chhe scenarios, zero se le kar aap ke phone par aik Personal AI Employee tak. Gateway-first harness, hands-on.
  • Hermes with General Agents: memory-first harness. Persistent context, self-improving skills, aik aisa agent jo aap ka kaam seekhta hai aur models ke aar paar portable rehta hai.

Is section ke aakhir tak aap aik aise worker ke maalik honge jo jab aap so rahe hon tab jawab deta hai, jo pichhle hafte aap ne use jo sikhaya woh yaad rakhta hai, aur aise infrastructure par chalta hai jo aap ki hai. Yahi farq hai AI use karne aur usay employ karne ke darmiyan.