Przejdź do treści
Biblioteka skilli
AI i agenci · Planowanie

agent-design-and-tooling

/agent-design-and-tooling

Projektuje pętlę agenta, narzędzia, uprawnienia, warunki zatrzymania i ślady z poziomów autonomii, zanim wybierze się framework.

Repozytorium
ai-pm-skills ↗
Licencja
MIT
Ostatnia zmiana repozytorium
13.11.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
40
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Projektuje pętlę agenta, narzędzia, uprawnienia, warunki zatrzymania i ślady z poziomów autonomii, zanim wybierze się framework.

Wyzwalacze z opisu

make it an agent
Oryginalny opis (EN)

Design an agent''s loop, tools, permissions, stop conditions and traces from its autonomy levels, before any framework. Use for "make it an agent", a new tool or permission, or a framework debate.

Jak działa

  1. When to use
  2. Inputs
  3. Method
  4. Output
  5. Rules and red lines
  6. Links
  7. Sources
Sekcje z nagłówków SKILL.md, w oryginale · 48 linii, 781 słów
Podgląd SKILL.md
---
name: agent-design-and-tooling
description: 'Design an agent''s loop, tools, permissions, stop conditions and traces from its autonomy levels, before any framework. Use for "make it an agent", a new tool or permission, or a framework debate.'
license: CC-BY-4.0
metadata:
  stage: "build"
  version: "0.3.1"
  status: "tested"
  eval_threshold: "0.80"
  autonomy_relevant: "true"
  depends_on: "llm-feature-prd, process-ai-redesign"
  feeds: "system-prompt-and-guardrails, assistant-kpis-cost-per-outcome"
  sources: "anthropic-agents, anthropic-agent-tools, mcp-spec, own-nanoagent, repo-chanmeng"
---

# agent-design-and-tooling

## When to use
- Someone asks to "make it an agent", or a workflow has branches and tools.
- A framework debate starts before the design exists.
- An agent gets a new tool, a higher cap or a new permission, and its matrix must change.
- Do not use for the prompt and guardrail wording: use `system-prompt-and-guardrails`.
- Do not use when one model call answers the need: say so and stop.
- Do not use to test the agent: that is `agent-evals`.

## Inputs
Mandatory: the process steps with their L-levels, the candidate tools.
Ask if missing: failure modes, latency and cost constraints.

## Method
1. Set the trigger (request, batch, event, schedule) and define done as a state of the world (a record written, a case ready for the adjuster), not a sentence in the transcript. Name the exits and what follows each: success, error, escalation, paused for approval, policy denied, budget exhausted. A trigger other than on-request also needs dedup of repeated signals.
2. Pick the loop type: single step, plan-execute, multi-agent. Start with a raw loop. Multi-agent needs a written justification naming what a single loop cannot do.
3. List tools with schemas, side effects and whether they can be undone. For each, write the contract the model reads: when to use it over its nearest neighbour, named parameters over a generic tool, and an empty result kept distinct from a failed call. Rate the risk by effect, not by name. Detail in references/tool-contracts.md.
4. Build the permission matrix: tool × level × approval, following references/autonomy-ladder.md. A tool that moves money or changes records carries a cap and an approval rule.
5. Write the protocol for each write tool at L3 and above: prepare is separate from commit; where the matrix requires approval, it is bound to the exact action, target and payload, and it expires; the cap is checked on the resulting state in the system of record, where other sessions and channels count too; every write has an idempotency key and is never retried blindly; a write that times out has an unknown result, so reconcile before any retry. How the gate is enforced and attacked belongs to `system-prompt-and-guardrails`.
6. Write stop conditions for the run. A run ends on a checkable done state or the API's stop signal, never on the model's prose or self-reported confidence. Maximum steps, budget per run and timeout are the safety net: a hit is an alert with its own exit (hand off to a human, never silent). Stop conditions for the initiative itself start in `ai-opportunity-assessment`; how the agent is tested over repeated trials is `agent-evals`.
7. Set the memory policy: what persists, for how long, and personal-data limits.
8. Define the trace schema: every tool call logged with input, output and cost, and an outcome field set from the exit, usable by `assistant-kpis-cost-per-outcome`.
9. Before handing the design to engineering, prototype the tool set on a local MCP server (synthetic or anonymised data, no production credentials: references/build-to-learn.md) and run about 10 representative tasks. Read the traces from step 8: which tools the agent called, in which order, where it looped or picked the wrong neighbour. Fix tool names, descriptions and response sizes before touching the prompt. The prototype is evidence for the design, not the build; record it in the Prototype check section, or "not yet run".
10. Map frameworks last: each candidate against a named need the raw loop cannot meet, or none.

## Output
`templates/agent-design.md`, all headings present. The loop line reads `Loop type: <type>`.
When the brief asks which framework to take, the answer still opens with what the design lacks, says the choice waits for it, and keeps candidates for the closing section. No conditional pick up front ("X, if you adopt one"): readers stop at the name.

## Rules and red lines
- Permissions follow the ladder; no write tool without a line in the matrix.
- No write tool at L3 or above without its protocol line.
- No agent without stop conditions and named exits.
- Multi-agent needs a written justification.
- The framework comes last, in the answer as much as in the method.

## Links
references/autonomy-ladder.md · references/tool-contracts.md · references/build-to-learn.md · system-prompt-and-guardrails · agent-evals · assistant-kpis-cost-per-outcome · process-ai-redesign

## Sources
- anthropic-agents: simple loops first, frameworks last.
- anthropic-agent-tools: prototype a few tools, run realistic tasks, read the transcripts, then refine.
- mcp-spec: tools exposed by a server to the model, which a local prototype server reuses.
- own-nanoagent: raw loop reference, trace with an outcome field.
- repo-chanmeng: ideas for tool contracts, the write protocol, trigger and exits, rewritten in our own words.
Pliki w paczce (40)
  • SKILL.md 5 KB
  • evals/agent-design-and-tooling-001/prompt.md 1 KB
  • evals/agent-design-and-tooling-001/fixtures/bad.md 1 KB
  • evals/agent-design-and-tooling-001/fixtures/good.md 1 KB
  • evals/agent-design-and-tooling-001/graders/empty-vs-failed.md 1 KB
  • evals/agent-design-and-tooling-001/graders/matrix.md 1 KB
  • evals/agent-design-and-tooling-001/graders/outcome.md 1 KB
  • evals/agent-design-and-tooling-001/graders/redline-no-multi.md 1 KB
  • evals/agent-design-and-tooling-001/graders/single.md 1 KB
  • evals/agent-design-and-tooling-001/graders/stop.md 1 KB
  • evals/agent-design-and-tooling-001/graders/triggered.md 1 KB
  • evals/agent-design-and-tooling-002/prompt.md 1 KB
  • evals/agent-design-and-tooling-002/fixtures/bad-no-approval.md 1 KB
  • evals/agent-design-and-tooling-002/fixtures/bad.md 1 KB
  • evals/agent-design-and-tooling-002/fixtures/good.md 1 KB
  • evals/agent-design-and-tooling-002/graders/approv.md 1 KB
  • evals/agent-design-and-tooling-002/graders/budget.md 1 KB
  • evals/agent-design-and-tooling-002/graders/cap.md 1 KB
  • evals/agent-design-and-tooling-002/graders/matrix.md 1 KB
  • evals/agent-design-and-tooling-002/graders/redline-pay-capped.md 1 KB
  • evals/agent-design-and-tooling-002/graders/triggered.md 1 KB
  • evals/agent-design-and-tooling-002/graders/write-timeout-unknown.md 1 KB
  • evals/agent-design-and-tooling-003/prompt.md 1 KB
  • evals/agent-design-and-tooling-003/fixtures/bad.md 1 KB
  • evals/agent-design-and-tooling-003/fixtures/good-run1.md 3 KB
  • evals/agent-design-and-tooling-003/fixtures/good-run3.md 3 KB
  • evals/agent-design-and-tooling-003/fixtures/good.md 1 KB
  • evals/agent-design-and-tooling-003/graders/budget.md 1 KB
  • evals/agent-design-and-tooling-003/graders/done-state.md 1 KB
  • evals/agent-design-and-tooling-003/graders/max-steps.md 1 KB
  • evals/agent-design-and-tooling-003/graders/raw.md 1 KB
  • evals/agent-design-and-tooling-003/graders/redline-framework-not-first.md 1 KB
  • evals/agent-design-and-tooling-003/graders/stop.md 1 KB
  • evals/agent-design-and-tooling-003/graders/triggered.md 1 KB
  • evals/agent-design-and-tooling-004/prompt.md 1 KB
  • evals/agent-design-and-tooling-004/graders/not-triggered.md 1 KB
  • references/autonomy-ladder.md 5 KB
  • references/build-to-learn.md 6 KB
  • references/tool-contracts.md 3 KB
  • templates/agent-design.md 1 KB

Wersje

W bazie nie ma innych skilli o nazwie agent-design-and-tooling.

Migawka bazy z 13.11.2026 · wersja paczki 13ae9cb5c20c (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

  • 100agent-spec-generator
    Pisze gotową do produkcji specyfikację agenta AI: zakres, narzędzia i uprawnienia, poziomy autonomii, reguły zatrzymania, eskalacja, tryby awarii, ewaluacje i wdrożenie.
  • 90agent-profile-design
    Przegląda lub projektuje konfigurację agenta kodującego: rolę, narzędzia, skille, kontekst i uprawnienia, bez tworzenia agenta.
  • 90agentic-workflow-design
    Projektuje wieloetapowe przepływy agentowe w Microsoft Foundry: kroki, progi autonomii, punkty kontrolne, budżety i warunki zatrzymania.
  • 95ai-agent-architectures
    Projektuje architekturę pojedynczego agenta AI: pętlę działania, narzędzia, pamięć, planowanie i ewaluację.
  • 100agent-tool-builder
    Definiuje narzędzia agentów we wzorcu fail-closed: jedna klasa z nazwą, schematem, bezpieczeństwem i wykonaniem oraz trójwarstwowe wykonanie walidacja, uprawnienia, wywołanie.
  • 90claude-agent-sdk-tool-integration
    Integruje narzędzia, uprawnienia i serwery MCP z agentami Claude przy użyciu Agent SDK.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.