Human-Agent Teams: Aap ki Workforce ka Operating Model

Yeh course AI workers ki us team ka operating model hai jo logon ke saath chalti hai. Is ki bunyadi unit ek qabil-e-aitemad worker hai, yani Digital FTE: woh loop chalata hai, searchable memory se kaam karta hai, apni identity se sign in karta hai, aur edges par escalate karta hai. Is track mein aap yeh worker banate hain; yeh operating model pehle paper par likh sakte hain, phir jab live workers online aa jayein to isay un se wire kar sakte hain. Ek qabil-e-aitemad worker unit hai. Un ki team chalana ek worker banane se mukhtalif skill hai, aur yeh course wahi hai: ek ko bohat mein badalne ka tareeqa.

Bohat se workers ki team, ek worker ka bara version nahin hoti. Woh mukhtalif cheez hoti hai, aur us ke liye mukhtalif skill chahiye: worker banana nahin, balkeh un workers ki team ko logon ke saath chalana.

Yeh course wahi operating model hai. Is ke baad ke chaar courses machinery hain: ek lead agent jo board hire karta hai (Workforce with Paperclip), ek workforce jo khud barhti hai (Self-Expanding Workforce), delegated approval (Identic AI), aur woh workers jo kama sakte hain (Payment-Enabled Agents). Yeh machinery us team par kaam nahin karti jise chalana aap ne seekha hi na ho. Is liye workforce automate karne se pehle aap yeh set karte hain ke humans aur workers ek roster, ek workspace, aur ek goal kaise share karte hain.

Yeh kis qisam ka course hai, is par ek note. Baqi workforce courses build-along hain. Yeh nahin. Yahan aap bohat kam code likhenge. Aap operating documents likhenge (roster, role cards, north star, verification rubric), bilkul us tarah jaisay manager likhta hai, bas farq yeh hai ke aap ka agent draft karega aur faisla aap karenge. Deliverables woh agreements hain jin par team chalti hai. Yeh code se kam glamorous hain, lekin zyada faisla-kun: zyada human-agent teams technology par nahin, practices par fail hoti hain.

Yeh section ka sab se accessible course bhi hai. Roles, goals, trust, aur kaun-kis-cheez-ka-owner-hai woh cheezen hain jo aap logon ke saath kaam kar ke pehle se samajhte hain. Agents in fundamentals ko nahin badalte. Woh inhein sahi karne ka daao barha dete hain.

Yeh practices kahan se aati hain

Yahan ke patterns Anthropic ke apne human-agent teams chalane ke account se liye gaye hain, aur is book ke pehle se banaye hue frameworks par map kiye gaye hain (poore links aakhir mein Sources mein hain). Jahan Anthropic koi specific result report karta hai, woh un ka hai aur unhi ke naam se diya gaya hai. Jin features par yeh depend karta hai (shared team tools mein kaam karne wale agents, apni credentials aur memory wale agents), wohi capabilities aap is track mein banate hain.

📚 Teaching Aid

Open Full Slideshow

Full presentation dekhein — Human-Agent Teams

Aap kya banayenge (artifact set)

App nahin: operating documents ka set jis par aap ki team chalti hai. Starter aapko har document template ke taur par deta hai; aap inhein apne agent ki madad se fill karte hain.

• Ek team roster: har member, human aur agent, role, owner, tools, aur autonomy level ke saath.
• Har agent ke liye ek role card: woh kya own karta hai, kya nahin karta, us ki tools, us ka kaam kaise check hota hai, aur woh kab escalate karta hai.
• Ek working agreement: default public kya hai, chand security boundaries kya hain, kya private rehta hai.
• Ek north-star doc: team ka ek ambitious goal, aur kaun se agents bina prompt ke us par act kar sakte hain.
• Ek verification rubric: work product kaise grade hota hai, taake har line human ke parhne ke baghair us par trust kiya ja sake.
• Ek doer-verifier setup: dusra agent jiska sirf ek kaam hai: pehle agent ko check karna.
• Ek weekly report: "lessons and missteps" log jo team ko behtar banata hai.
• Ek attention budget: aap kya review karenge, kya batch hoga, aur aap tak pohanchne wali cheezon ki cap.

Setup

1. Starter download karein (human-agent-teams-starter.zip) aur unzip karein. Yeh templates ka folder hai, code nahin. Inhein kisi bhi editor mein kholein.
2. Behtar hai ke ek Digital FTE (Building a Digital FTE) ho jiske gird aap real team chalayen. Abhi worker nahin? Theek hai: yeh course planning mode mein karein (neeche note), phir manual ko live worker se wire karein jab woh ban jaye.
3. Aisi jagah ho jahan kaam team ko nazar aata ho: shared channel, doc library, repository. Agents wahi parhi hui cheezon se seekhte hain.
4. Apna agent draft karne ke liye tayyar rakhein (claude.ai, Cowork, ya aap ka worker). Starter ka har artifact isi rhythm se fill hota hai: aap direct karte hain, agent draft karta hai, aap faisla karte hain.

Yahan se har Part ek practice sikhata hai, phir aap se woh document likhwata hai jo us practice ko zameen par rakhta hai. Aap theory par quiz de kar nahin niklenge; aap ek team ka operating manual le kar niklenge.

Readiness check (Part 2 se pehle yeh karein)

Yeh course assume karta hai ke aap ka worker pehle se aap ki team ka written record parh sakta hai. Abhi test karein: apne agent se kahen ke pichle haftay ka koi decision ya document dhoondhe, aise channel mein jo us ka own nahin. Agar woh kar sakta hai, aap ready hain. Agar khaali wapas aaye, to aap ne AI Searchable Context wala searchable system of record abhi complete nahin kiya. Pehle woh karein. Us ke baghair yahan ki har practice ke paas parhne ko kuch nahin.

Abhi wahan tak nahin pohanche? Isay planning mode mein chalayein

Technical stack se pehle bhi aap yeh poora course kar sakte hain: claude.ai ya Cowork ko drafting agent banayein, tamam operating documents likhein, aur har agent role ko "active" ke bajaye "planned" mark karein. Aap paper par complete operating manual ke saath niklenge. Jab pehle workers ban jayein, wapas aakar planned roles ko live roles se swap kar dein.

---

Part 1: Ek worker se team tak

Concept 1: Single-player khatam ho chuka hai

AI ke saath kaam pehle single-player tha: ek shakhs, ek chat window, ek task. Digital FTE is se pehle hi zyada karta hai. Is course ka shift multiplayer hai: kai log aur kai agents ek workspace mein, shared goals ki taraf kheenchte hue. Humans strategy set karte hain; agents execute karte hain.

Multiplayer agent woh hai jo ek waqt mein kai humans ke saath kaam karta hai. Digital FTE ki tarah us ki apni memory aur skills hoti hain. Chat window ke baraks, us ke paas apni credentials hoti hain (kisi person se borrowed nahin), aur woh jahan kaam hota hai wahin rehta hai: team ke channels aur docs mein, private session mein nahin.

Unit Digital FTE hai. Team humans aur Digital FTEs ka ek shared roster hai. Team hi business hai.

Concept 2: Worker ko kin parts ki zaroorat hoti hai

Team tab tak kaam nahin karti jab tak har agent ke paas teen cheezen na hon, aur yeh track teeno banata hai:

• Persistent memory: taake woh goal ko dinon tak pakar kar rakhe, sirf ek prompt tak nahin (AI Searchable Context).
• Apni identity: credentials jo human se tied na hon, taake woh kisi ke logins borrow karne ke bajaye aap ke set kiye hue guardrails ke andar act kare (AI Identity).
• Broad, searchable access: taake woh likhi hui cheezon se seekhe ke organisation kaam kaise karti hai (aap ka Postgres system of record aur RAG: retrieval, woh searchable memory jo aap ne usay di).

In ke baghair, "team mein agent add karna" ka matlab hota hai ek person apna password script ke saath share kar raha hai. In ke saath, iska matlab roster ka hissa banne wala worker hai. Aap operating model ab design kar sakte hain aur jab yeh teen parts live hon to isay live workers se wire kar sakte hain; human practices dono surat mein upar baithti hain.

✓ Checkpoint: aap unit ko jaante hain. Memory, identity, aur access wala worker woh cheez hai jisse team banti hai. Ab aap in mein se bohat se workers ko logon ke saath kaam karwate hain.

Concept 3: Scarce resource human judgment hai

Poora operating model ek cheez bachata hai: human attention aur judgment. Agents tez aur bohat hain; log bottleneck aur authority hain. Is course ki har practice is liye hai ke humans sirf woh decisions karein jo sirf humans ko karne chahiye, aur baqi sab se bahar rahen.

Pehle failure mode ka naam rakhein, kyun ke yeh aam hai. Operating model ke baghair log side par personal AIs ki fleets chalate hain. Kaam duplicate hota hai. Team ka context private windows mein toot jata hai jise koi aur, human ya agent, nahin dekh sakta. Jis metric ki sab ko zaroorat hoti hai woh paanch mukhtalif tareeqon se compute hota hai. Fix zyada agents nahin; fix ek team ko open mein chalana hai.

Baqi course chaar practices hai jo yahi kaam karti hain.

✓ Checkpoint: aap shape jaante hain. Chaar practices, ek maqsad. Ab pehli practice.

---

Part 2: Open mein kaam

Concept 4: Agar likha nahin, to woh mojood nahin

Agent apni samajh puri tarah un cheezon se banata hai jise team searchable banati hai: channels, code, docs, notes. Private messages, hallway conversations, aur restricted files us tak nahin pohanchte. Agent ke liye jo likha nahin gaya woh nazar hi nahin aata.

Is liye pehli practice technical hone se pehle cultural hai: public mein kaam karein. Decisions channels aur docs mein land karte hain, direct messages aur be-notes meetings mein nahin. Artifacts aise likhein ke agent unhein dhoondh sake: agent ab aap ki documentation ka primary reader hai, afterthought nahin.

Payoff real hai, aur Anthropic isay seedha report karta hai. Jo agent team ke decisions parh sakta hai woh aisa kaam pitch nahin karega jise aap pehle hi kill kar chuke hain. Jo dusri team ke specs parh sakta hai woh kaam kar chuka pattern reuse karega. Aur kyun ke agent kisi bhi human se bohat tez parhta hai, woh aksar relevant kaam surface karta hai jo log miss kar dete. Transparency virtue se leverage ban jati hai.

Concept 5: Boundaries workspace par, document par nahin

Agent kya dekh sakta hai yeh decide karne ka ghalat tareeqa hai: ek document, ek channel, har dafa. Yeh humans aur agents dono ke liye decision fatigue hai: kya yeh private ho? kya main woh doc share kar sakta hun? kya yeh agent us thread mein allowed hai? Soft, per-item lines thakane wali hoti hain aur unhein ghalat karna aasan hota hai.

Sahi tareeqa: chand clear security boundaries jo workspace level par draw hon: security boundary sirf information ke set ke gird deewar hai, aur ek rule ke andar kaun hai. Boundary ke andar context har teammate, human ya AI, tak flow karta hai. Clear lines ki choti tadaad soft lines ki bari tadaad se behtar hai, aur yeh roz ka "kya main yeh share kar sakta hun?" tax hata deti hai.

Yahin aap ka system of record apni jagah kamata hai. Boundary deewar hai; AI Searchable Context ka searchable store woh cheez hai jo us ke andar azaadi se flow karti hai. Deewar ek dafa draw karein; retrieval ko baqi kaam karne dein.

Exception seedhi kahiye, kyun ke public-by-default ka matlab everything-is-public nahin. Kuch kaam sensitive hota hai aur ek human aur ek agent ke beech rehna chahiye. Woh agent ko direct message hai, ya private apps (claude.ai, Cowork) aap ke personal connectors par, jahan conversation private rehti hai. Default open rakhein; jo waqai private rehna chahiye us ke liye clear, narrow lane rakhein.

Draft it. 01-working-agreement.md kholein aur apne agent mein paste karein:

Draft a working agreement for my team. State what is public by default. List the few security boundaries we need (no more than a handful) and who is inside each. List what stays private (one human, one agent). For each boundary, write one sentence a new teammate could follow.

Check it. Kya aap har boundary ko ek single sentence mein keh sakte hain? Agar nahin, boundaries zyada hain. Few and clear, warna yeh tikti nahin.

✓ Checkpoint: context flow karta hai. Aap ki team wahan kaam karti hai jahan agents parh sakte hain, chand aisi walls ke peeche jinka naam koi bhi le sakta hai. Ab kaam ko names dein.

---

Part 3: Ek roster, clear roles

Concept 6: Team ka roster hota hai

Human-agent team ek roster, ek artifact set, aur ek working space share karti hai. Is liye roster likhein: har member, human aur agent, aur har ek kya own karta hai.

Agents mukhtalif roles rakhte hain. Ek data analysis own karta hai; ek design standard hold aur enforce karta hai; ek research synthesis chalata hai. Jab project shuru hota hai, humans agents se baat karte hain ke kaun se roles assign karne hain aur mil kar kaise kaam karna hai: roster us conversation ka output hai, pehle se lagaya gaya guess nahin.

Yeh aap ki Roles Taxonomy aur Digital FTE taxonomy hai, ek team ke liye concrete banai hui. Catalog batata hai ke workers ki kaun si qisam mojood ho sakti hai; roster batata hai ke is team par kaun hain aur kaun kya own karta hai.

Concept 7: Role ek card bhi hai, aur skill file bhi

Har agent ko ek role card milti hai: woh kya own karta hai, kya nahin own karta, kaun se tools aur access chahiye, us ka kaam kaise check hota hai, aur woh kab human ko escalate karta hai. Scope "does not own" ke baare mein utna hi hai jitna "owns" ke baare mein: fuzzy edges wala agent dusron ke kaam mein drift karta hai.

Tools ka naam likhein, kyun ke un ke baghair role sirf be-hath title hai. Analyst ko database chahiye. QA agent ko browser tool chahiye. Har role ke access list karein, aur sirf wahi grant karein (least privilege wohi rule hai jo delegated approval mein phir milega).

Phir role ko skill file ke taur par likhein. Yeh woh move hai jo book ke frameworks ko click karwata hai: agent ka role skill mein define karein, aur role portable ban jata hai: org mein koi bhi us se usi type ka dusra agent khara kar sakta hai. Roles org chart ke boxes rehna band kar dete hain aur copy ki ja sakne wali skills ban jate hain. (Skills is poori book ka portable lever hain; role ek aur cheez hai jo skill carry kar sakti hai.)

Human-only roles ko explicit rakhein. Humans unhi threads mein kaam karte hain jahan agents karte hain, lekin woh roles humans ke paas rehte hain jo sirf humans rakh sakte hain: consequential calls, cost wala judgment. Roster se aap human judgment ko un decisions par rakhte hain jahan zaroorat ho, aur un se hata dete hain jahan zaroorat nahin.

Jab agent ko dusre agent ki zaroorat ho

Kabhi job ek worker ke liye bohat bari hoti hai, aur lead agent sub-task ke liye sahi context ke saath teammates spawn karta hai: yahan researcher, wahan reviewer. Yeh instinct sahi hai, aur next course isi ko automate karta hai: Workforce with Paperclip "lead hires a board" ko budgets aur approvals ke andar managed workforce bana deta hai. Aap ka roster aur role cards us ke inputs hain. Yahan aap roles haath se likhte hain taake samajh sakein ke Paperclip baad mein aap ke liye kya karega.

Underlying feature par mid-2026 tak do honest notes: Claude Code agent teams experimental hain aur default se disabled hain (aap setting se on karte hain), aur sirf lead teammates spawn karta hai; teammates apne nested agents nahin bana sakte. Is liye "agents spinning up agents" asal mein "lead ek flat team spawn karta hai" hai. Isay early samjhein, aur production mein depend karne se pehle current docs parh lein.

Draft it. 02-roster.md aur 03-role-cards/role-card.template.md ki copy kholein aur paste karein:

Draft a team roster for [team]. List every member, human and agent. For each: role, who owns it, the tools and access it needs, and its autonomy level. Mark the roles only a human should hold. Then write a full role card for [my worker]: owns, does NOT own, tools/access, how its work is verified, and what triggers an escalation to a human.

Check it. Har member ka owner aur "does not own" hai. Har agent ke paas tools aur ek clear escalation trigger hai. Agar do members same task claim kar sakte hain, scopes abhi sharp nahin.

✓ Checkpoint: har kisi ka lane hai. Humans aur agents ek roster par, har ek named job own karta hai aur us ke tools rakhta hai. Ab team ko direction dein.

---

Part 4: North star

Concept 8: Aisa goal jo agent ko proactive banata hai

Context aur roles agent se woh kaam karwate hain jo aap assign karte hain. North star us se sahi kaam propose karwati hai. North star ek ambitious, wide-reaching goal hai jo team ko batata hai ke kaun se tasks aur workstreams karne ke qabil hain: woh ek sentence jiske against baqi sab measure hota hai. Humans hamesha isay set karte hain, business mission mein grounded.

Jab yeh likh diya jaye, aap isay team ke agents ke saath share karte hain. Phir (aur yahi part log skip kar dete hain) aap naam le kar batate hain ke kaun se agents bina prompt ke is par act kar sakte hain. Har agent ko work propose nahin karna chahiye. Sirf woh agents jinke paas skills aur earned trust hai.

Anthropic ki example choti aur exact hai: ek team jiska north star tha "product onboarding ko zyada helpful banana" us ke agent ne proactively onboarding error messages rewrite karne ki recommendation di: aisi changes jinhon ne next week onboarding success measurably barhai. Agent ne poochne ka wait nahin kiya. North star ne usay bataya ke rewrite on-mission hai.

Yeh aap ki AI-Native Company mission hai, ek team tak push down ki hui. Company ki mission hai; team ki north star hai jo us mission ko serve karti hai; agent ka kaam hai jo north star ko serve karta hai. Line goal se task tak seedhi chalti hai.

Concept 9: Proactivity woh privilege hai jo aap grant karte hain

Proactive agent ka risk yeh hai ke agent woh kaam propose kare jise usay chhoona nahin chahiye. Is liye proactivity named hoti hai, assumed nahin. Aap kehte hain kaun se agents workstreams suggest kar sakte hain, aur north star woh test hai jo har proposal ko pass karna hota hai. Jis agent ke paas yeh grant nahin, woh phir bhi apna assigned job karta hai: bas freelance nahin karta.

Draft it. 04-north-star.md kholein aur paste karein:

Help me write a north star for [team]. It should be one ambitious goal, grounded in our mission. State why it matters. Name which agents on the roster may propose new work against it, and the guardrails on those proposals. Write it so an agent, given only this doc, could judge whether a new idea is on-mission.

Check it. Isay named agent ki nazar se parhein. Sirf is doc ke saath kya woh on-mission idea ko off-mission idea se alag kar sakta hai? Agar nahin, star steering ke liye zyada vague hai.

✓ Checkpoint: team ke paas direction hai. Ek goal, humans ka set kiya hua, chand named agents ko usay chase karne ki ijazat. Ab decide karein ke aap unhein kitna run karne dete hain.

---

Part 5: Trust, earned

Concept 10: Autonomy reliability ke saath barhti hai

Aap naye colleague ko pehle din keys nahin dete. Aap agent ko bhi pehle din 500 bug fixes nahin dete. Anthropic ke engineers wahan tak pohanche (agents hundreds of fixes apne aap handle karne lage), lekin shuruat wahan se nahin hui. Autonomy demonstrated reliability ke proportion mein grant karein, phir har task type ke liye deliberatively widen karein.

Kisi task ko achi tarah karne ka tacit knowledge externalise karne ke liye feedback cycles lagte hain: naye human ke liye bhi, agent ke liye bhi. Aur models badlein to retest karein: weaker model ko madad dene wali guardrail stronger model ko shackled kar sakti hai, aur model improve hone par prompt ko reword karna par sakta hai. Trust ek dafa set nahin hota; tune hota hai.

Ladder ko operational banane ke liye fixed rungs dein. Roster mein autonomy level har agent per task type set karein, poore agent ke liye ek level nahin:

Level	What the agent does	Where the human is
L0	Drafts only; the human does the work	human does everything
L1	Acts, but a human reviews every output	human reviews all
L2	Acts; a verifier checks; human reviews only exceptions	human reviews exceptions
L3	Acts within limits; batches escalations to the human	human reviews batched escalations
L4	Runs the task type on its own, within approved scope	human reviews the weekly report

Naya agent kisi task type par L1 se shuru karta hai aur repeated, verified wins ke baad upar jata hai. Wahi agent ek task type par L4 aur dusre par L1 ho sakta hai: autonomy worker-on-a-job ko grant hoti hai, worker ko general taur par nahin.

Concept 11: Kaam ko checkable banayein

Jo cheez autonomy ko safely grow karne deti hai woh yeh hai: kaam human ke dekhne se pehle verify ho sakta hai. Code ke paas tests hote hain, zahir hai. Lekin zyada tar dusra kaam bhi grade ho sakta hai: document rubric aur style guide ke against, report checklist ke against. Jab aap bar set karte hain aur har assignment ko vettable banate hain, quality high rehti hai aur aap ke intended direction se drift nahin karti.

Yeh team level par Eval-Driven Development hai (Eval-Driven Development). Wahan eval worker ko automatically grade karta hai. Yahan rubric wohi eval hai jo ek worker ke output par apply hota hai: wahi idea, checklist ki shakal mein jo teammate chala sakta hai.

Phir doer-verifier: ek agent task karta hai, dusre agent ka sirf ek kaam hota hai usay check karna. (Anthropic isay doer-verifier harness kehta hai.) Yeh sasti insurance hai, aur yeh human ka waqt bachane ke liye agent ka waqt kharch karti hai: verifier drift ko aap ki scarce attention kharch hone se pehle pakar leta hai.

Draft it. 05-verification-rubric.md aur 06-doer-verifier.md kholein aur paste karein:

Write a verification rubric for [my worker]'s main output: the concrete checks that decide whether the work is good enough to ship, in plain pass/fail terms. Then describe a doer-verifier setup: a second agent whose only job is to grade the first's output against this rubric and return pass/fail with reasons.

Check it. Kya second agent sirf is rubric ko use kar ke pehle agent ka kaam grade kar sakta hai, aur kya aap us pass par trust karenge? Agar "pass" ke baad bhi aap har line parhna chahte hain, rubric specific nahin.

Concept 12: Human attention ko paisay ki tarah kharch karein

Jab agents independent hote hain, naya failure mode aata hai: humans output mein doob jate hain. Is liye human attention ko scarce resource samjhein. Behtareen teams apne agents se questions batch karwati hain ek single pass mein, key context repeat karwati hain taake human jaldi up to speed ho, aur items ki tadaad limit karti hain jo human ek waqt mein dekhta hai.

Kuch teams ek agent ko sirf yeh kaam deti hain ke humans tak kya elevate karna hai. Kuch cap lagati hain ke agent har din kitna kaam kare: is liye nahin ke usay slow karna hai, balkeh is liye ke humans ab bhi kaam ke saath meaningful tareeqe se engage kar sakein, aur un skills ko zinda rakhein jo un ke liye important hain.

Cycle mein reflection build karein. Team se weekly "lessons and missteps" report mangain, taake mistakes track hon aur repeat hona band karein. Track karein ke har agent ne kaun se task types par autonomy earn ki hai, aur scope sirf repeated wins ke baad widen karein. Report se team qismat ke bajaye irade se behtar hoti hai.

Draft it. 07-weekly-report.md aur 08-attention-budget.md kholein aur paste karein:

Draft a weekly team report template that captures, for each agent: what it shipped, its lessons and missteps this week, and which task types it has earned more autonomy on. Then propose an attention budget for me: what I will review, what gets batched, and the cap on how much reaches me at once.

Check it. Busy week mein kya yeh human ko sirf important cheezon ka faisla karne deta hai, aur kuch nahin? Agar human ko ab bhi sab kuch parhna parta hai, budget scarce resource protect nahin kar raha.

✓ Checkpoint: trust switch nahin, dial hai. Kaam checkable hai, autonomy proof ke saath widen hoti hai, aur human attention wahan kharch hoti hai jahan woh count karti hai. Aap ke paas poora operating model hai.

---

Part 6: Apni team khari karein

Aap ne chaar practices seekhi hain aur har ek ke liye document draft kiya hai. Ab inhein ek team ke operating manual mein assemble karein.

Operating manual: ek folder, aath files

Manual ek folder hai, us order mein numbered jismein aap isay fill karte hain. Starter exactly yeh ship karta hai:

human-agent-team/
  01-working-agreement.md      few clear boundaries · what's public · what's private
  02-roster.md                 every member · owner · tools · autonomy level (L0–L4)
  03-role-cards/               one card per agent (copy the template)
    role-card.template.md
    reconciler.md              (filled example)
  04-north-star.md             the one goal · which agents may act on it unprompted
  05-verification-rubric.md    the pass/fail checks a verifier can apply
  06-doer-verifier.md          which agent checks which, and what happens on fail
  07-weekly-report.md          shipped · lessons & missteps · autonomy changes
  08-attention-budget.md       what you review · what's batched · the cap

Har file ki choti required checklist hoti hai (template mein bhi, aur har Part ke end par "Check it" ke taur par repeat bhi). File tab tak done nahin jab tak us ki checklist all yes na ho. Manual tab tak done nahin jab tak tamam aath done na hon.

Isay order mein fill karein

Order dependency order hai. Chaar practices paanch fill-steps mein map hoti hain (trust practice verification aur attention mein split hoti hai), aur yeh aath files banati hain: ek manual teen zoom levels par.

1. Working agreement: kya public hai, chand boundaries, kya private rehta hai. (Context pehle; is ke baghair kuch aur kaam nahin karta.)
2. Roster + role cards: har member, woh kya own karta hai, us ke tools, us ke escalation triggers.
3. North star: goal, aur kaun bina prompt ke usay chase kar sakta hai.
4. Verification rubric + doer-verifier: aap ke dekhne se pehle kaam kaise check hota hai.
5. Weekly report + attention budget: team kaise improve karti hai aur aap ka waqt kaise protect karti hai.

Har ek ko usi rhythm se chalayein: Part ka prompt paste karein, agent ka draft parhein, aur faisla karein: cut, sharpen, approve. Authority aap hain; agent drafter hai.

Anthropic ke paanch sawalon ko done-test banayein. Team tab ready hai jab har jawab yes ho:

1. Kya agents aur humans ko chahiye information aur access dono public aur broadly searchable hain?
2. Kya aap apni team ka roster, humans aur agents, likh sakte hain aur keh sakte hain har member kya own karta hai?
3. Kya har human aur agent ke paas apna job karne ke liye sahi tools hain?
4. Kya key work products verify karne ke liye rubrics ya tests hain?
5. Kya team ke paas clear north star hai jise sab reference kar sakte hain?

Worked example: finance close team

Templates abstract rehti hain jab tak aap filled example na dekhein. Yeh choti finance team hai jo monthly close chalati hai (ek human controller aur teen agents), jahan important parts concrete hain. (Starter isay examples/finance-close-team.md ke taur par ship karta hai.)

North star: building se bahar jane wala har number sahi ho aur apne source tak traceable ho.

Member	Human/Agent	Owns	Tools / access	Autonomy
Controller	Human	Sign-off on anything that leaves the company	none	human-only
Puller	Agent	Pulling figures from the source systems	ERP / GL read-only	L2 (verified)
Reconciler	Agent	Matching figures across sources, flagging variances	the ledger, the system of record	L3 on routine ties; L1 on new accounts
Checker	Agent	Grading the reconciliation against the rubric	the rubric	doer-verifier only

Jo detail isay safe banati hai woh escalation trigger hai, Reconciler ki role card par seedha likha hua.

Example: Reconciler ka escalation trigger

Controller ko escalate karein jab: koi variance account balance ke 1% ya $10,000 se zyada ho, jo bhi chota ho (deliberately conservative, taake choti accounts bhi choti swings par escalate hon), ya koi figure system of record mein source ke baghair ho. Warna usay tie karein aur log karein.

Aur verification rubric jo Checker apply karta hai. Reconciliation sirf tab pass hoti hai jab:

Example: Checker ki rubric

1. every balance ties to its source within threshold; 2. every variance has a reason code; 3. every source document is linked in the system of record; 4. every exception is listed in the escalation queue.

Woh escalation line miniature mein poora operating model hai. Reconciler routine ties apne aap chalata hai (L3), Checker kisi ke dekhne se pehle rubric ke against verify karta hai (doer-verifier), unsourced ya material numbers ruk kar human tak pohanchte hain (attention sirf wahan kharch hoti hai jahan count karti hai), aur Controller woh ek role hold karta hai jo bahar ki duniya ko number ship karta hai. Note karein Reconciler routine ties par L3 lekin new accounts par L1 hai: autonomy per task type, per agent nahin. Thresholds aur sources swap karein, aur yahi shape accounts payable, payroll, ya board reporting chalata hai.

✓ Checkpoint: aap team chala sakte hain. Working agreement, clear roles wala roster, north star, kaam verify karne ka tareeqa, aur apni attention ka budget. Yeh operating model hai, aur baqi workforce courses isi par chalte hain.

---

Part 7: Ceiling, jahan yeh grow karta hai

Operating model akela team ko scale nahin karta. Yeh rules set karta hai; agle chaar courses woh machinery hain jo in rules par chalti hai, aur har ek aap ke likhe hue artifact ko input banata hai:

• Workforce with Paperclip roster automate karta hai: lead agent budgets, approvals, aur full audit trail ke andar workers ka board hire aur run karta hai. Aap ka roster aur role cards wahi hain jin se woh hire karta hai.
• Self-Expanding Workforce kaam ke barhne ke saath team ko grow karti hai, bajaye is ke ke aap har worker haath se add karein.
• Identic AI aap ka attention budget automate karta hai: signed identity jo aap ki set ki hui limits ke andar routine approvals clear karti hai aur sirf consequential ones surface karti hai.
• Payment-Enabled Agents worker ko transact karne deta hai: cost bachane wali team se kamaane wali team tak ka qadam.

Pehle operating model banayein, phir us machinery ke paas chalne ke liye mazboot cheez hogi. Isay skip karein, aur aap aisi team automate kar rahe honge jo shuru se coherent hi nahin thi.

Aur practices ka ceiling khud: humans ke liye is mein kuch naya nahin. Clear north star, defined roles, open mein kaam, quality ke liye shared bar, mistakes se seekhne ki jagah: yeh healthy team habits hain jo hum decades se jaante hain. Agents inhein introduce nahin karte. Woh inhein skip karna fatal bana dete hain, kyun ke agent buri practice ko bhi achi practice ki tarah tez scale karega. Jo teams agents se sab se zyada hasil kar rahi hain woh fundamentals par sab se zyada disciplined hain.

Yeh woh line hai jahan tak book chal kar aa rahi thi: Digital FTEs ki workforce, is operating model par chalti hui, ek AI-native company ke andar. Aap yahan ek worker ke baare mein soch kar aate hain. Aap yahan se nikalte hain to logon ke saath un ki team chala sakte hain, aur us team ke output ko scale, govern, aur sell kar sakte hain.

Wahi manual, dusri teams

Artifact set ek hi shape hai; team badalti hai, documents nahin:

• Research team: analyst, synthesiser, aur fact-checker agents, "answer the question, with sources" north star ke neeche.
• Delivery team: planner, doer, aur doer-verifier, quality rubric ke neeche, ship decision human hold karta hai.
• Finance team: data-pull agent, reconciliation agent, aur woh human jo building se bahar jane wale har number ka owner hai.

Wahi paanch documents. Mukhtalif roster, mukhtalif north star, mukhtalif rubric.

Capstone: real team khari karein

Apni organisation ka real goal chunen aur us ke liye full artifact set produce karein: working agreement, roster, role cards, north star, verification rubric, doer-verifier, weekly report, attention budget.

1Your Work

Apna operating manual (ya us ki key files) paste karein: working agreement boundaries, roster table, ek ya do role cards apne escalation triggers ke saath, north star, verification rubric, aur attention budget.

2Get Your Score

Discuss with an AI. Question your scores.
Come back when you have your BEST evaluation.

Starter graded example (examples/finance-close-team-graded.md) ship karta hai, ek complete finance manual jo in aath checks ke against 15/16 score karta hai, ek weak check named aur fix shown ke saath. Apna manual grade karne se pehle isay parhein: yeh dikhata hai ke rubric kya pakarta hai aur strong manual kaisa lagta hai.

Sources

Yeh course Anthropic ke human-agent teams chalane ke account se sikhata hai, aur is book ke pehle se banaye hue frameworks par map karta hai. Primary source aur Anthropic material jisse yeh draw karta hai:

• Anthropic, "Lessons from Anthropic on building effective human-agent teams" (June 2026): primary source. Chaar practices, aur text mein named specific results (onboarding error-message rewrite, 500-bug-fix trust arc, workspace-level security boundaries), yahan se hain.
• "Equipping agents for the real world with Agent Skills" aur Agent Skills overview: agent ke role ko portable skill file ke taur par define karna.
• "Managed agents" aur agent memory: humans se tied na hone wali credentials, aur persistent memory (Concept 2).
• "Effective context engineering for AI agents": broad, searchable context (Part 2).
• "Harness design for long-running agents": doer-verifier harness (Concept 11).
• Agent teams in Claude Code aur Claude Tag: shared team spaces mein apni identity ke saath kaam karte agents.

Flashcards Study Aid

---

Apni samajh test karein

Checking access...