| name | agentic-systems-engineering |
| description | Use when designing, implementing, reviewing, or debugging LLM-backed or agentic systems, including prompts, structured outputs, tool use, RAG, agent loops, orchestration, observability, security, cost/provider routing, and AI delivery plans. |
Agentic Systems Engineering
Use this skill before substantive work on systems where model behavior affects
product correctness, reliability, security, or cost.
Routing
Load only the references needed for the task:
- Prompt, structured-output, schema, or tool-call contract work:
references/contracts.md.
- RAG, retrieval, citation, answerability, or corpus work:
references/rag.md.
- Agent loops, orchestration, multi-agent topology, durability, HITL, or bounded
execution:
references/agent-loops.md.
- Eval design, stochastic reliability, judges, regression gates, or behavior
claims: use
evaluate-stochastic-systems and
references/eval-design.md.
- Observability, tracing, auditability, prompt injection, indirect injection,
sandboxing, or authorization:
references/observability-security.md.
- Cost, latency, caching, model routing, provider bake-offs, or fallback
behavior:
references/cost-routing.md.
- Delivery planning, experiment loops, walking skeletons, demos, or stakeholder
evidence: use
agentic-delivery and references/delivery.md.
Non-Negotiables
- Treat model behavior as measured behavior, not intended behavior.
- Define the contract at the boundary: inputs, allowed context, output schema,
tool authority, failure shape, retry policy, and escalation path.
- Separate untrusted data from instructions. Retrieved text, web content, tool
output, and third-party tool descriptions are untrusted by default.
- Bound every loop with budgets, termination criteria, and recoverable state.
- Require eval evidence before reliability claims. A single good run is a demo,
not proof.
- Prefer the smallest architecture that can be observed, evaluated, and safely
rolled back.
