agent-engineer
/agent-engineer
Wskazuje, który skill agentowy pasuje do projektu LLM, od prostego agenta czy chatbota po przegląd całej bazy kodu.
- Autor
- sarthakrastogi ↗
- Repozytorium
- agent-engineer ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Wskazuje, który skill agentowy pasuje do projektu LLM, od prostego agenta czy chatbota po przegląd całej bazy kodu.
Wyzwalacze z opisu
Oryginalny opis (EN)
Entry point for LLM agent work when the right agent-* skill isn't obvious. Maps symptoms to the specialised agent-* skills and defines the optional agent-engineering/ project record. Use when starting a blank-page agent, chatbot or RAG project ("build an agent that…", "where do I start?"), for a broad request spanning several areas ("we're launching next week", "review our agent"), or to survey an existing LLM codebase before changing it. For a specific task (a prompt, tools, evals, tracing, retrieval, cost) load that skill directly instead.
Jak działa
- Where is it failing? (the agent stack)
- Shortcuts: the user already has something
- Evidence
- Starting in an existing codebase
- Project record (optional)
- Delegation
Podgląd SKILL.md
---
name: agent-engineer
description: >-
Entry point for LLM agent work when the right agent-* skill isn't obvious. Maps symptoms
to the specialised agent-* skills and defines the optional agent-engineering/ project
record. Use when starting a blank-page agent, chatbot or RAG project ("build an agent
that…", "where do I start?"), for a broad request spanning several areas ("we're
launching next week", "review our agent"), or to survey an existing LLM codebase before
changing it. For a specific task (a prompt, tools, evals, tracing, retrieval, cost) load
that skill directly instead.
license: MIT
metadata:
version: 0.4.0
---
# Agent engineer
Pick the specialist skill and get out of the way. Each `agent-*` skill carries its own
rules; this one only maps problems to skills and holds the shared conventions.
## Where is it failing? (the agent stack)
Each layer contains the one below: graph ⊃ loop ⊃ harness ⊃ context ⊃ prompt. Fix at the
lowest layer that explains the failure; add a layer only on evidence.
| Symptom | Layer | Go to |
|---|---|---|
| Right facts in view; wrong wording, format or tone | Prompt | `agent-prompting` |
| Model didn't see the right facts or tools | Context | `agent-context`, `agent-rag` |
| Unsafe action, no sandbox or verification, lost state | Harness | `agent-tools`, `agent-guardrails` |
| Runs forever, stops early, retries blindly | Loop | `agent-design`, `agent-production` |
| Roles collide or handoffs lose work | Graph | `agent-design` (most tasks never need one) |
Wrong answers or hallucinations with no obvious layer → `agent-accuracy` localises them.
New agent → `agent-design`, then tools → prompting → observability → evals.
About to launch → `agent-production` → `assets/launch-checklist.md`.
Salesforce Agentforce (`.agent` files, Agent Script, `sf agent`) → `agentforce`, which maps
these skills onto the platform.
## Shortcuts: the user already has something
| Situation | Do this |
|---|---|
| Two versions + an eval set | `agent-evals` → *Comparing two versions* |
| One objective failure (invented IDs, broken format) | `agent-evals` → *Minimal eval*, then `agent-accuracy` |
| No tracing, need to see what it does | `agent-observability` → *Debug-grade tracing* |
| Inherited eval suite | `agent-evals` → *Auditing an existing suite* |
An existing agent without evals or tracing gets the minimum the task needs. Name the gap;
don't block on it.
## Evidence
For a non-trivial behaviour change, propose the smallest eval that would show it worked.
If the user declines, or the change is cosmetic, make it and say plainly that it's
unevaluated. Never call an unmeasured change an improvement: mocked or predicted results
and offline unit tests show the code runs, not that the agent got better.
Report in your reply or the PR, not in new summary or report files the user didn't ask
for.
## Starting in an existing codebase
1. Find the LLM calls, prompts, tool definitions and agent loop.
2. Check for tracing, eval sets and logged production data.
3. Read `agent-engineering/` if present. It holds past decisions.
4. Summarise in ≤ 10 lines: architecture, models, tools, what's measured, biggest gap.
Propose the next step.
## Project record (optional)
For an agent project that spans sessions, offer once to keep `agent-engineering/` (or
`$AE_DIR`) in the repo; create it only if the user agrees. If it exists, keep it current.
Otherwise put decisions and eval deltas in your reply or the PR description.
| File | Template | Holds |
|---|---|---|
| `design.md` | `agent-design` → `assets/design.md` | success criteria, architecture, risks, production plan, decision log |
| `eval-plan.md` | `agent-evals` → `assets/eval-plan.md` | evals, graders, judge validation, gates |
| `failure-taxonomy.md` | `agent-evals` → `assets/failure-taxonomy.md` | failure modes with counts |
| `experiment-log.md` | `agent-accuracy` → `assets/experiment-log.md` | per behaviour change: hypothesis, eval delta, decision |
| `observability.md` | `agent-observability` → `assets/observability.md` | span schema, metrics, PII policy |
| `datasets/*.jsonl` | | eval sets, retrieval and attack sets |
## Delegation
If subagents are available: `trace-analyst` for error analysis over many traces,
`eval-engineer` for datasets and judges, `rag-diagnostician` for retrieval failures,
`agent-reviewer` for a pre-release review. They can't ask the user anything, so put what
they need in the brief, including whether the user agreed to `agent-engineering/` files.
Wersje
W bazie nie ma innych skilli o nazwie agent-engineer.
Migawka bazy z 13.11.2026 · wersja paczki 55af974e8162 (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
- 100building-ai-agents
Projektuje, buduje i przegląda agentów AI na sprawdzonym stacku: wybór modelu, RAG, MCP, pamięć, bezpieczeństwo i bramki wydania. - 100llm-app-engineering
Buduje niezawodne funkcje oparte na LLM: prompty, dane wyjściowe, RAG, narzędzia, ewaluacje, koszty i bezpieczeństwo. - 100ai-engineer
Buduje funkcje produktowe oparte na LLM jako inżynier aplikacji AI: RAG, workflow agentowe, prompty, wywoływanie narzędzi, ustrukturyzowane wyjście, ewaluacje i guardrails. - 90production-ai-agents
Podpowiada, jak budować agenta LLM, RAG i funkcje z wywoływaniem narzędzi: model, prompty, kontekst, koszty tokenów i ewaluacje. - 100senior-prompt-engineer
Optymalizuje prompty i szablony, ocenia odpowiedzi LLM, projektuje RAG i architektury agentów oraz analizuje zużycie tokenów. - 90prompt-engineering
Tworzy, poprawia i przegląda prompty do LLM, takie jak prompt systemowy, skill albo agent, oraz diagnozuje ich błędy.