Vy tol'ko chto zapustili vtorogo agenta Claude Code. Teper' u vas problema.

Pervyy agent v seredine refaktoringa. Vtoromu nuzhno postroit' funktsiyu, kotoraya zatragivaet nekotorye iz tekh zhe faylov. Ni odin ne znaet o sushchestvovanii drugogo. Vy odnovremenno router, khranilishche sostoyaniya i razreshitel' konfliktov, i vash edinstvennyy instrument -- kopirovanie konteksta mezhdu oknami terminala.

Imenno zdes' bol'shinstvo razrabotchikov upirayutsya v stenu s Claude Code. Ne potomu chto agent plokho kodit. A potomu chto net sistemy, kotoraya govoriala by emu, nad chem rabotat'.

Problema kopirovaniya-vstavki

Bol'shinstvo rabochikh protsessov s Claude Code nachinayutsya odinakovo. U vas est' zadacha v golove (ili v Jira, ili v GitHub issue), i vy vstavlyaete opisanie v prompt agenta. "Postroy potok autentifikatsii." "Isprav' bag paginatsii." "Dobav' podderzhku temnoy temy."

Dlya odnogo agenta eto rabotaet normal'no. U agenta polnyy kontekst, vy mozhete nablyudat' za vyvodom, i vy znaete kogda on zakonchil, potomu chto vy na nego smotrite.

Dobav'te vtorogo agenta -- i treshchiny poyavlyayutsya mgnovenno.

Agent A refaktorit sloy API. Agent B stroit novyy endpoint. Oba zatragivayut server/routes.ts. Ni odin ne znaet ob izmeneniyakh drugogo. Vy obnaruzhivaete konflikt, kogda odin delaet push i rabota drugogo lomaetsya. Ili khuzhe: oba uspeshny lokal'no, no ob"yedinennyy rezul'tat sloman sposobami, kotorye ni odin diff ne pokazyvaet.

Kornevaya prichina ne v tom, chto agenty neakkuratny. Eto otsutstvie obshchego sostoyaniya. Net mesta, gde "Agent A otvechaet za refaktoring API" bylo by zafiksirovano. Net statusa, kotoryy govorit "fayl marshrutov redaktiruetsya, zhdi svoyu ochered'." Agenty rabotayut na otdel'nykh promptakh bez kakogo-libo ponimaniya obshchey kartiny.

Dobav'te tret'ego agenta -- i vy tratite bol'she vremeni na koordinatsiyu, chem na kod.

Chto agentam deystvitel'no nuzhno ot sistemy zadach

Prezhde chem iskat' instrument, stoit sprosit': chto na samom dele nuzhno agentu Claude Code dlya khoroshey raboty?

Udivitel'no malo.

Unikal'nyy identifikator. Chto-to, na chto on mozhet ssylat'sya v kommitakh i kommentariyakh. "Ispravil bag" bespolezno v mul'tiagentonom loge. "Zavershen PROJ-47: paginatsiya vozvrashchaet nepravil'nyy podschet na filtrovannykh vidakh" -- eto otslezhivaemo.

Chetko opredelennyy ob"em. Zagolovok, opisanie i kriterii priemki. Ne roman. Ne pol'zovatel'skaya istoriya s personami. Konkretnoe opisanie togo, kak vyglyadit "gotovo." "Endpoint /users vozvrashchaet stranichnyye rezul'taty. Razmer stranitsy po umolchaniyu 25. Pole next_cursor ravno null na posledney stranitse."

Status, kotoryy mozhno obnovlyat'. Agent dolzhen signalizirovat', gde on nakhoditsya: zayavlen, v protsesse, zavershon. Bez etogo vy snova zaglyadyvaete v okna terminala i gadaete.

Osvedomlennost' o zavisimostyakh. "Ne nachinay eto poka PROJ-46 ne zamerdzhen" predotvrashchaet samyy rasprostranennyy sboi mul'tiagenta: stroitel'stvo na kode, kotoryy eshcho ne sushchestvuet.

Obratite vnimanie na to, chto otsutstvuet v etom spiske. Planirovanie sprintov. Otslezhivanie skorosti. Kanban-doski. Story points. Epiki s tsvetnymi metkami. Agentam ne nuzhen teatr upravleniya proektami. Im nuzhna zadacha, status i sposob skazat' "ya zakonchil."

Kontrakt CLAUDE.md

Sistema zadach govorit agentam, nad chem rabotat'. Fayl CLAUDE.md govorit im, kak rabotat'.

Esli vy zapuskaete neskol'ko agentov Claude Code, u kazhdogo dolzhen byt' CLAUDE.md, opredelyayushchiy ego identichnost' i granitsy. Eto ne optsional'naya nastroyka. Eto raznitsa mezhdu agentami, kotorye koordiniruyutsya, i agentami, kotorye meshayut drug drugu.

Vot uproshchennyy primer dlya inzhenernogo agenta:

## Identity

Engineer for the project. You implement features, fix bugs,
and write tests. You own implementation quality.

## What You Own

- All files under `components/` and `lib/`
- Unit tests in `__tests__/`
- You may read but not modify files under `server/`

## What You Don't Own

- Deployment configuration (that's ops)
- Issue triage and prioritization (that's the coordinator)
- QA validation (QA tests your work independently)

## Completion Protocol

Before marking any task done:
1. Run the full test suite: `pnpm test`
2. Verify your change works manually
3. Comment what you did with the commit hash
4. Push before reporting completion

Razdel granits -- eto nesushchaya chast'. Bez yavnogo vladeniya faylami agenty otklonyayutsya. Inzhenernyy agent "usluzhlivo" refaktorit konfiguratsiyu deploya. Agent QA "ispravlyaet" test, izmenyaya testiruyemyy kod vmesto samogo testa. Yavnye granitsy predotvrashchayut eti rezhimy sboya.

Protokol zaversheniya vazhen ne men'she. On predotvrashchaet samyy rasprostranennyy sboi agenta: zayavlyat' chto-to gotovym, kogda ono vsego lish' kompiliruetsya. "Zapusti polnuyu test-syuitu" i "prover' vruchnuyu" -- eto konkretnye vorota. Agent, sleduyushchiy etomu protokolu, sozdaet rabotu, kotoroy chelovek mozhet doveryat'. Agent bez nego sozdaet rabotu, kotoruyu nuzhno proveryat' postrochno.

Masshtabiruy'te eto na neskol'kikh agentov -- i poluchite flot, gde kazhdyy chlen znaet svoyu polosu, svoy protokol peredachi i chto oznachaet "gotovo."

CLI-first upravlenie zadachami

Vot nablyudenie po rabochemu protsessu, kotoroye zanyalo u menya bol'she vremeni, chem sleduet: agenty Claude Code rabotayut dramaticheski luchshe s CLI-instrumentami, chem s GUI-interfeys ami.

Eto logichno, esli podumat'. Agent Claude Code zhivet v terminale. On mozhet vypolnyat' komandy, chitat' vyvod i predprinimat' deystviya na osnove rezul'tatov. Prosit' ego navigirovat' veb-interfeys, nazhi mat' knopki i interpretirovat' otrendirovannyye stranitsy -- eto bor'ba s yestestvennym interfesom agenta.

Sistema zadach na baze CLI oznachaet, chto agent mozhet sdelat' vsyo eto v odnom potoke:

# Read the task
task show PROJ-47

# Claim it
task update PROJ-47 --status in_progress --assignee agent-1

# Do the work...

# Report completion
task comment PROJ-47 "DONE: Fixed pagination. Commit: abc1234"
task update PROJ-47 --status done

Nikakogo pereklyucheniya konteksta. Nikakikh okon brauzera. Nikakikh skrinshotov kanban-doski. Agent chitaet zadachu, vypolnyaet rabotu i obnovlyaet status -- vsyo ne pokidaya sredy, v kotoroy on rabotaet.

Vyvod takzhe machinoschitaem. Kogda nuzhno proverit', chto proiskhodit mezhdu agentami, mozhno zaprosit':

task list --status in_progress    # What's being worked on?
task list --assignee agent-2      # What is agent-2 doing?
task list --blocked               # What's stuck?

Eto i est' forma instrumentariya, kotoryy rabotaet. CLI, kotoryy govorit na yazyke agenta.

Beads: lokal'nyy issue-treker dlya agentov

Rabochiy protsess, opisannyy vyshe, ne gipotetichesk iy. Eto to, chto ya ispol'zuyu kazhdyy den' s beads, open-source lokal'nym issue-trekerom, postroennym imenno dlya takoy razrabotki, upravlyaemoy agentami.

beads khranit issue (nazyvaemyye "beads") v lokal'noy baze dannykh Dolt ryadom s vashey kodovoy bazoy. Kazhdyy bead imeet ID, zagolovok, opisanie, status, prioritet, zavisimosti i potok kommentariyev. CLI nazyvaetsya bd, i eto interfeys, cherez kotoryy agenty chitayut zadachi, obnovlyayut statusy i ostavlyayut strukturirovannyye kommentarii.

Vot real'nyy rabochiy protsess. Ya sozdayu zadachu:

bd create --title "Fix pagination on filtered views" \
  --description "The /users endpoint returns wrong count when filters are applied. Page size defaults to 25. next_cursor should be null on the last page." \
  --priority p2

Agent zayavlyaet yeyo:

bd update bb-r3k2 --claim --actor eng1
bd update bb-r3k2 --status in_progress

Prezhde chem pisat' kod, agent kommentiruet svoy plan:

bd comments add bb-r3k2 --author eng1 "PLAN:
1. Fix count query in /users to apply filters before COUNT()
2. Add cursor boundary check for last page
3. Add test cases for filtered pagination

Files:
- server/routes/users.ts - fix count query
- server/routes/users.test.ts - add filtered pagination tests"

Eto kontrolnaya tochka. Esli plan oshibochen, vy lovite eto za 30 sekund vmesto togo, chtoby obnaruzhit' plokhuyu realizatsiyu 45 minut spustya.

Agent vypolnyaet rabotu, zapuskaet testy i kommentiruet zavershenie:

bd comments add bb-r3k2 --author eng1 "DONE: Fixed filtered pagination count.

- COUNT() now applies the same WHERE clause as the data query
- next_cursor returns null when offset + page_size >= total_count
- Added 4 test cases covering filtered + unfiltered pagination

Commit: a1b2c3d"

bd update bb-r3k2 --status ready_for_qa

Zadacha teper' imeet polnyy audit-treyl: chto bylo zaprosheno, chto agent zaplaniroval, chto on fakticheski sdelal, i khesh kommita dlya revyu. Vtoroy agent, vypolnyayushchiy QA, mozhet podkhvatit' yeyo i proverit' nezavisimo.

Eto rabotaet potomu chto beads govorit na tom zhe yazyke, chto i agenty. Vsyo -- eto CLI-komanda. Vsyo proizvodit strukturirovannyy vyvod. Net nesootvetstviya impedansa mezhdu instrumentom i agentom.

Uvidet' obshchuyu kartinu

CLI-rabochiy protsess masshtabiruetsya do 3-4 agentov, prezhde chem vy upreetes' v novyy potolok. Ne instrumental'nyy potolok. Kognitivnyy.

Pri 5 agentakh zapusk bd list i mental'naya sborka sostoyaniya proekta -- eto kak chitat' tablitsu i pytatsya uderzhat' graf zavisimostey v golove. Kakiye zadachi zablokirovany? Kakoy agent ne obnovlyal svoy status 20 minut? Epik funktsii gotov na 60% ili 80%? Informatsiya vsya tam v vyvode CLI, no yeyo sborka trebuet usiliy, kotoryye nakapliv ayutsya s kazhdym dopolnitel'nym agentom.

Vot gde vkhodit Beadbox. Eto dashboard real'nogo vremeni, kotoryy raspolagaetsya poverh beads i pokazyvaet sostoyaniye vsey vashey floty agentov. Vizual'no otrendirovannye derev'ya zavisimostey. Progressbary epikov. Potoki kommentariyev agentov, kotoryye vy mozhete prosmatrivat' bez zapuska pyati komand bd show.

Beadbox ne zamenyaet CLI. Agenty po-prezhnemu ispol'zuyut bd dlya vsego. Beadbox -- eto sloy, kotoryy vy otkryvaete, kogda vam nuzhna obshchaya kartina: kakiye rabochiye potoki dvizhuyutsya, kakiye zastryali, i gde uzkie mesta. On otslezhivaet bazu dannykh beads na predmet izmeneniy i obnovlyaetsya v real'nom vremeni, tak chto vy nikogda ne smotrite na ustarevshuyu informatsiyu.

Besplatno vo vremya bety i rabotaet polnost'yu na vashey mashine. Nikakikh akkauntov, nikakogo oblaka, vashi dannyye ostayutsya lokal'nymi.

S chego nachat'

Vam ne nuzhny 13 agentov, chtoby poluchit' pol'zu ot strukturirovannogo upravleniya zadachami. Nachnite s dvukh agentov Claude Code i odnogo pravila: kazhdaya zadacha poluchaet bead, kazhdyy agent kommentiruet svoy plan pered kodom, kazhdoye zavershenie vklyuchaet shagi verifikatsii.

Pattern nakapl ivaetsya. Kak tol'ko u agentov poyavlyaetsya obshchaya sistema zadach, vy mozhete dobavit' QA-agentov, kotoryye verif itsiruyut rabotu nezavisimo. Mozhete dobavit' koordinatora, kotoryy raspredelyaet zadachi iz ochered i prioritetov. Mozhete masshtabirovat'sya do 5, 10, 15 agentov bez lineynogo rosta nakladnykh raskhodov na koordinatsiyu, potomu chto protokoly berut na sebya to, chto ran'she bylo ruchnym pereklyucheniem konteksta.

Instrumenty:

beads dlya lokal'nogo upravleniya zadachami. Open source.
Claude Code kak sreda vypolneniya agentov.
Beadbox dlya vizual'nogo nablyudeniya, kogda flot rastet.

Esli vy stroyite rabochiye protsessy vrode etogo, postav'te zvezdu Beadbox na GitHub.

Kak upravlyat' zadachami dlya agentov Claude Code