Design Agents
Architecture before scaffolding: tools, memory, evals, stop conditions.
What it is
Section titled “What it is”Architecture before scaffolding: tools, memory, evals, stop conditions. This practice is a skill-mapped TTP: when, why, and which command—not a full copy of the skill. Open the skill to execute.
Why it works
Section titled “Why it works”This is regime C design (Quality regimes): quality is eval + trace shaped, not “assert exact string.” Effective agent systems stay simple — clear tools, explicit handoffs, and evaluation — not maximal swarms (Anthropic). Design-time limits on agency reduce OWASP “excessive agency” and sensitive-data leakage. Bake in datasets, graders, and observability (house default: Langfuse on OTel). Treat max-steps and deny-lists like non-functional requirements.
When to use it
Section titled “When to use it”When the situation matches the one-liner above and Orientation (or your project path) says this is the fire to touch now.
- Invoke the skill; follow its safety rules
- Keep one write owner; collect evidence before claiming done
- Hand off to the next practice instead of boiling the ocean
- Skip orientation when you’re lost
- Spawn overlapping agents to “go faster”
- Treat the practice as done without evidence
Related concepts
Section titled “Related concepts”05-agent-agency, 07-stop-conditions, 11-quality-regimes
Further reading
Section titled “Further reading”- Quality regimes (handbook)
- Anthropic — Demystifying evals for AI agents
- Anthropic — Building effective agents
- Langfuse docs
- OWASP Top 10 for LLM Applications
See also the handbook Sources & grounding bibliography.
Agent skill
Section titled “Agent skill”agents design