team-coordination
/team-coordination
Porządkuje pracę wielu agentów w zespole Agents Hub: role, przekazania, rozmowy na kartach, wspólną pamięć i zużycie tokenów.
- Autor
- mynamezxc ↗
- Repozytorium
- hermes-agent-hub ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 05.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 05.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Porządkuje pracę wielu agentów w zespole Agents Hub: role, przekazania, rozmowy na kartach, wspólną pamięć i zużycie tokenów.
Oryginalny opis (EN)
Use when work needs more than one agent in an Agents Hub team, or when you pick up a Kanban card: roles, hand-off, talking on cards, shared memory and files, unblocking, token economy.
Jak działa
- 1. Know the team and where things are
- 2. Hand work off: one Kanban card per outcome
- 3. While you work a card: be visible
- 4. When you are stuck: ask first, block last
- 5. Finish: hand over and learn
- 6. Follow up and accept (for the creator)
- 7. Token economy (why the rules above)
- 8. Report to the owner
Podgląd SKILL.md
---
name: team-coordination
description: "Use when work needs more than one agent in an Agents Hub team, or when you pick up a Kanban card: roles, hand-off, talking on cards, shared memory and files, unblocking, token economy."
version: 2.0.0
---
# Working as an agent team
Each member is a real agent. It keeps its own identity (`SOUL.md`), its own memory and skills, and it builds its
own experience across tasks. The team works like co-workers on a shared board: everyone can see who is doing what
and how far they are, they talk on the work item itself, and they share one small team memory and a folder of files.
Nobody needs the whole context pasted into every message.
Use other members only when the work really needs them. If doing it yourself is faster, do it yourself.
## 1. Know the team and where things are
- `<root HERMES_HOME>/agents-hub/TEAM.md`: id, name, role, who they report to, model, strengths. The owner edits the
roster in the Hub. Hand work down to people who report to you, report up to the one you report to.
- `<root HERMES_HOME>/agents-hub/shared/`: the **shared team space**.
- `NOTES.md` is team memory: one dated line per durable fact (decision, convention, where something lives, who
owns what).
- `files/` holds files handed between members.
- Read `NOTES.md` when you start a card (it is short on purpose).
- Your **own memory** is your experience: lessons, pitfalls, the owner's preferences for your kind of work, where your
inputs live. The memory tool writes it into your profile and it loads into every one of your sessions.
## 2. Hand work off: one Kanban card per outcome
Hermes's dispatcher starts the assignee as a full worker under its own profile, soul, memory, skills, permissions and
approvals. Create cards with `kanban_create`:
- `assignee` (id from TEAM.md), `title` (the outcome, short), `parents` (cards that must finish first).
- **`body`: only what is specific to this card.** Goal, inputs (paths), output (path), acceptance criteria,
deadline, who reviews.
- Shared rules belong in the assignee's SOUL/memory or one rules file you point to (`Read <path> first`).
- Do not paste the same rules into every card. Long repeated bodies cost tokens on every run and still get skimmed.
- Grant what the work needs **in the card**, explicitly. Example: "may write files in `<output dir>`", "may log in to the
local test site with <how>". A worker that has to guess its permissions stops and asks.
- Never start another agent yourself with `hermes ... chat --oneshot`; for a small sub-task with no separate role, use
`delegate_task`.
## 3. While you work a card: be visible
- `kanban_heartbeat` **with a short note** at each milestone ("3/8 articles written", "tests running"). The Hub's
Workflow tab and the delegator's chat panel show the note as your progress; a heartbeat without a note shows nothing.
- Talk on the card with `kanban_comment`: questions to the card's creator, findings a reviewer needs, a hand-over note.
Mention the other agent by id (`@qa`). Comments are the team's conversation, and the Hub lists them per card.
- Put large outputs in files (`shared/files/` or the card's output path) and **comment the path**, not the content.
- Use `kanban_show` only for what you need. The card body plus your memory should be enough to start.
## 4. When you are stuck: ask first, block last
1. Look for the answer in the card, its comments, `shared/NOTES.md`, your memory, and the rules file the card points to.
2. If an agent can answer (usually the card's creator), ask it on the card with a comment. Continue with the parts
you can do.
3. Block (`kanban_block`) only when nothing else can move. Pick the right kind:
- `needs_input`: the creator or the owner must decide.
- `capability`: a hard wall only the owner can remove (access, money, credentials).
- `dependency`: waiting on another card.
4. A block reason says what you need, from whom, and what is already done. The Hub shows it as "Waiting on …".
5. Never block on something already granted in the card, its rules file or your SOUL. Re-read before blocking.
Creators: check your cards that wait on you (Hub → Overview "Team now", Tasks → Blocked), answer on the card,
then `kanban_unblock`.
## 5. Finish: hand over and learn
- `kanban_complete` with a one-paragraph summary, `artifacts` (absolute paths) and machine-readable metadata.
- **Save experience before you finish**, 1–3 short items, only durable ones:
- your **memory**: a pitfall, a working method, an owner preference for your kind of work;
- your **skill**: a procedure you will repeat;
- `shared/NOTES.md`: a fact the whole team needs, as one dated line with your id and the path or card.
- The reviewer of your work is another member. When your card comes back with comments, fix exactly those.
## 6. Follow up and accept (for the creator)
- **Give several cards at once and wait on a condition.** Create the cards for your reports, then create ONE card for
yourself with `parents` = those cards (e.g. "Merge API + UI and report to manager"). It stays in todo and starts
by itself when every parent is done. That is how a superior "waits for A and B, then continues"; never poll.
- If you did not plan a follow-up, the Hub task heartbeat gives you a card "Nghiệm thu N thẻ cấp dưới…" when your
reports finish. Review the real output, comment "nghiệm thu" or create a fix card, plan the next step, then complete
it with a report your own superior can read. **Start the summary with `ĐẠT` or `KHÔNG ĐẠT`**: the Hub reads that
first sentence and shows "Accepted / Not accepted by <you>" on every card you reviewed.
- A report blocked on a question for you gets a card "Trả lời @…". Answer on the blocked card, `kanban_unblock` it,
then complete the answer card. What only the owner can decide (money, publishing, secrets, new access): say so on
the card and leave it waiting on the owner. The Hub lists it under "Waiting on you".
- Track each hand-off as one of your goals. Mark it `done` only after you checked the real output (files, tests,
logs). Do not trust self-reports.
- Not good enough: comment specific fixes or create a follow-up card. Still failing after two rounds: tell the owner.
## 7. Token economy (why the rules above)
- Your system prompt (soul, memory, skills index) is the same in every session, so the model provider caches it.
Do not change it mid-session. Keep memory entries short and few.
- Per-task context comes from the card and the files it points to. Read files on demand, the parts you need, instead
of loading everything.
- Hand summaries and paths to each other, not transcripts.
## 8. Report to the owner
Result, evidence, blockers, what needs approval, and who did what. Anything that needs the owner's decision (cost,
publishing, access, deleting data) is always asked first, never done on your own.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · mynamezxc/hermes-agent-hub | 100 | MIT |
| bybren-llc/safe-agentic-workflow | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki fd9ace2fbd9b (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100team-agent-orchestration
Organizuje pracę zespołu agentów przez elementy pracy, przypisanie, tablicę Kanban, bramki scalania i przekazania. - 95clawteam
Koordynuje roje wielu agentów przez CLI ClawTeam: zespoły agentów równolegle, zadania z zależnościami, komunikaty, tablica kanban i gotowe szablony zespołów na git worktree i tmux. - 90team-memory
Daje zespołowi agentów wspólną pamięć długoterminową, zapisuje poprawki i decyzje z podpisem autora i nie kopiuje danych osobowych. - 90agent-hub
Pomaga współpracować z innymi agentami kodującymi przez agent-hub, delegując zadania z pełnym kontekstem i zgłaszając wyniki. - 100team-agent-orchestration
Prowadzi orkiestrację zespołów agentów: zadania z właścicielami i zakresem, tablica Kanban, izolacja gałęzi, widok panelu kontrolnego i bramki scalania wyników w jeden produkt. - 90orchestrator
Kieruje pracę z wielu ról IT i gamedev do zespołu studia: liderów i wyspecjalizowanych agentów.