AgentPatterns

AgentPatterns: Product Playbook for an Engineering Knowledge Base

You want a knowledge base that engineers actually reuse: patterns, runbooks, and evaluable playbooks for agents/LLM systems—so teams ship faster without repeating the same failures.

intro

What this is (and what it is not)

AgentPatterns is a productized playbook repository for building and operating LLM/agent systems.

It is designed to be:

  • Actionable: copy/pasteable procedures and checklists
  • Operational: includes failure modes and rollback/diagnosis notes
  • Evaluable: “what good looks like” + minimal regression tests

It is not a blog. If it can’t be executed or verified, it doesn’t belong here.

Content contract: every playbook must answer

  1. When should I use this? (scope, prerequisites, constraints)
  2. What decision does it help me make? (tradeoffs, default recommendation)
  3. How do I implement it? (steps)
  4. How does it fail in production? (failure modes + observability)
  5. How do I know it works? (minimal eval + regression cases)

If a doc doesn’t have at least (2), (4), and (5), it is likely “advice”, not a playbook.

Publishing model

This site intentionally uses an admin-only publishing flow.

  • It keeps content consistent (no drive‑by edits).
  • It allows stronger opinions and clearer defaults.
  • It avoids building a full auth product before there’s traction.

The public site only lists playbooks with status = PUBLISHED.

Recommended first 10 playbooks (high leverage)

If you want this repository to become useful quickly, start with these:

  1. Tool‑Calling judgment: when to call tools vs. answer directly
  2. Tool‑call parsing strategy: safe fallbacks + audit logging
  3. RAG debugging: retrieval failures, wrong retrieval, context pollution
  4. Minimal eval harness: 20 cases, scoring rubric, regression runner
  5. Prompt+policy structure for agents (separation of concerns)
  6. Retry/backoff/circuit breaking for flaky tools
  7. Idempotency + dedupe for side-effect tools
  8. Observability: tracing, tool-call logs, error taxonomy
  9. Caching strategy: what to cache and what not to cache
  10. Red‑team prompts for your top 3 failure modes

References