A pattern language for agents

The practice

I'm Roman Kagan. I build production AI systems — Claude Skills, tools, and agents — and I help teams architect the ones they're building. The work spans search and retrieval engineering, agentic workflows, evaluation and guardrails, and the unglamorous infrastructure that lets any of it run safely in production.

The catalog on this site is the by-product of that practice: a structured account of the patterns I keep reaching for, written down so they can be named, compared, and reused.

The methodology

The series is consciously modeled on Martin Fowler's Patterns of Enterprise Application Architecture (Addison-Wesley, 2002). The first volume tested a single proposition — whether the Fowler-style catalog format could be applied to agentic AI architecture. The answer was yes, and the format produced enough value that sixteen more volumes followed.

Every pattern entry follows the same template:

• Intent — what the pattern accomplishes, in a sentence or two.
• Motivating Problem — the situation that requires it.
• How It Works — the mechanics, broken into clear steps.
• When to Use It — best-fit conditions and explicit limits.
• Sources — where the pattern is documented in industry practice.
• Example artifacts — prompts, agent code, and SKILL.md sketches where they apply.

The consistency is the point. As Fowler put it, a pattern language is about vocabulary: if a colleague says “we'll use Prompt Chaining for ingestion and an Orchestrator–Workers loop for the merge, with a Human-in-the-Loop Checkpoint before commit,” every architecturally important decision has been communicated in one sentence.

What this is — and isn't

It is reference material: a working library organized for retrieval, not for linear reading. It is not a textbook, not vendor documentation, and not an academic survey. The gap it fills is a Fowler-style pattern catalog applied to agentic AI as the discipline stands in 2026 — provisional by design, and built for revision as practice matures.

Citation

The content is authored by Roman Kagan (RelevantSearch.AI / Softfinity / Grid Dynamics) and is currently draft v0.1 reference material — not yet a published work. To cite a pattern, reference its volume and name, e.g. Kagan, R. “Orchestrator–Workers,” Patterns of AI Agent Workflows (Agentic AI Series, Vol. 1), 2026.