ai-patterns.com · reading guide
A series on moving multi-agent systems from fragile prompts to production-grade architecture — two maturity models, a 55-pattern reference catalog, installable Claude skills, and interactive tools for auditing your own system. Start wherever your question is.
The enterprise adoption ladder, Levels 0–6: data foundation → model serving → RAG → tuning → grounding & evaluation → single-agent → multi-agent. The ordering is a dependency chain — jump to Level 6 and you inherit every unresolved problem below it, now acting autonomously. Includes the Sense → Reason → Plan → Act loop and a horizontal-vs-vertical use-case explorer.
The other maturity model — Agentic Governance Maturity, six levels from Prompting to Collaborative Swarm. Each level adds a capability and a new class of failure; you mature by adding the control that contains it. Where the Blueprint asks "what should we build next," this asks "what must we contain before we can trust it."
All 55 patterns from the book, across seven categories and chapters 5–11, each with Context, Problem, Solution, Consequences and Source plus code (34 ship validated implementations). Robustness & Fault Tolerance is the largest category at 16 — the honest signal that the hard engineering is failure, not cooperation.
The same 55 patterns packaged as installable Claude Agent Skills — one directory each, with
SKILL.md, the full book treatment, and a reference implementation. Descriptions are written as
problem statements, so a skill fires when you describe the situation even if you never name
the pattern. Drop into ~/.claude/skills/.
Reasoning as observation at a chosen resolution — self-similar Macro / Meso / Micro / Nano apertures over one substrate. Makes "deeper" measurable via J⁺/J⁻ rubrics and a hill-climb gate (ΔJ⁺ ≥ +1.5, ΔJ⁻ ≤ −1), so relabelling fails visibly instead of passing as progress. Includes a copy-paste system prompt implementation.
Ten patterns as complete, copy-paste system prompt contracts, each with an architecture diagram and a runnable simulator that traces the agent loop step by step. The interesting runs are the refusals — a compliance monitor blocking a PHI request, an auditor reverting a "successful" discount — because a library that only shows happy paths teaches nothing about governance.
Where the research weight sits across all 55 patterns, what patterned architecture buys over raw prompting, and a 12-point production readiness audit that scores your system live — Foundational, Production-Ready, or Autonomous Fabric. Score it honestly; the controls people skip are exactly the ones that stay invisible until an incident.
A short definitional primer on FCoT. For the full pattern treatment — mechanisms, comparison against CoT/ReAct/ToT/Reflexion, liabilities, and an implementation — see the deep dive above.
An earlier pattern catalog focused on governance of agentic systems — the precursor to the governance maturity treatment in The Age of Emergent AI.