Agentic Design (Pattern Catalog)
A free, searchable architecture catalog for AI builders by KORTEXYA SAS: 280+ agent design patterns, 280+ techniques, 900+ use cases, and 280+ interactive demos. A connected catalog — not a list of buzzwords — that lets engineers compare, inspect, and apply design patterns for building reliable AI agents. The freemium model ("start with the library, pay when the tools save you work") adds a Prompt Optimizer, Eval Lab, and red-team audits in the Pro tier.
Key Quotes
"A connected catalog, not a list of buzzwords."
The site's self-description, and the thing that distinguishes it from the flood of AI reference sites that are really just SEO landing pages. The claim is that patterns are connected — you can trace from a constraint (reliability, latency, cost, safety) through relevant patterns to implementation options. This is the A Pattern Language (Christopher Alexander) model applied to agent architecture: each pattern names a recurring problem and its proven solution, and patterns compose.
"Start with the library. Pay when the tools save you work."
The freemium thesis in one sentence. The entire catalog is browsable for free; you pay for the Pro tools (Prompt Optimizer, Eval Lab, red-team audits) that operationalize the patterns. This is a smarter pitch than most AI tooling — it bets that the patterns themselves are marketing for the tools, and that engineers who learn the catalog will pay to automate what they now understand.
"Design AI agents with patterns you can inspect, compare and apply."
The verb choices matter: inspect (not just read), compare (side-by-side trade-offs, not isolated descriptions), apply (practical, not academic). This is engineering documentation written for builders, not researchers.
"Choose how an agent should reason, use tools and fail safely."
The production framing: agent design isn't about capability, it's about reliability. The hard part isn't making an agent that works; it's making one that fails gracefully in production. This aligns with Building Agents That Don't Break Themselves and the broader harness-engineering thesis.
Key Themes
The Pattern Language Revival
Christopher Alexander's 1977 A Pattern Language gave architects 253 composable patterns for buildings. Agentic Design applies the same model to AI agent architecture — and it's one of the first to do so at scale. The 280+ patterns aren't just a list; they're connected, so you can navigate from a concrete constraint to the relevant design patterns to implementation techniques. This is what distinguishes a catalog from a wiki: the connections are curated, not just hyperlinked.
4-Step Design Workflow
The site's workflow — Frame → Compare → Inspect → Practice — is a useful mental model even without the product. Frame forces you to name the constraint (reliability, latency, cost, or safety) before reaching for a pattern. Compare puts patterns side by side on mechanism, trade-offs, and use cases — the missing feature in most architecture documentation. Inspect dives into implementation with interactive examples. Practice closes the loop with learning paths and labs. This is a better design process than most teams follow for any architecture decision.
The Freemium Bet
The free tier is genuinely useful — full catalog access, interactive demos, an AI expert, sandboxed code execution. The Pro tier adds operational tools: Prompt Optimizer, Eval Lab, red-team audits. This is a calculated bet that (a) the catalog is valuable enough to attract a large free audience, (b) a fraction of that audience will pay for tools that automate what the catalog teaches, and (c) the catalog itself improves as more people use it. The bring-your-own-key option (Anthropic/OpenRouter) is smart — it avoids the margin-crushing token markup model.
What's Missing
The site claims 280+ patterns but the homepage doesn't name a single one. There's no sample pattern, no taxonomy overview, no sense of what "pattern" means in this context. Is a pattern a code template? A design decision tree? An architecture diagram? The 4-step workflow is clear, but the catalog's actual shape is invisible from the landing page. This might be a deliberate free-tier tease, or it might be that the patterns are thinner than the count suggests. Either way, you can't evaluate a pattern catalog without seeing a pattern.
Where It Fits
The agent design space is filling in. Elements of Agentic Systems Design gives you the ten-element taxonomy. The Agentic Product Standard v2.0 gives you the production standard and Claude Code skills to apply it. Agent-Native Architectures (Every) gives you the five principles for building applications around agents. Agentic Design gives you the pattern catalog — the connective tissue between these frameworks, organized by constraint rather than taxonomy. It's the reference manual in a stack that already has the textbook and the field guide.
A useful comparison: Software Engineering Practice Atlas takes the opposite approach — AI-generated breadth (4,654 cards across all of software craft) versus human-curated depth (280+ patterns in one domain). Together they form a natural experiment in whether curation or scale produces better patterns. Agent-Building Playbook (Ramparte) sits at a third point on the axis: open and git-native — 113 patterns with a validated flat-frontmatter contract — rather than a proprietary web catalog, optimising for machine-ingestibility and PR-driven contribution over interactive search.
The KORTEXYA Question
The site is produced by KORTEXYA SAS, a French company with no other visible products. Is this a startup building toward a paid platform? A consultancy's lead-gen tool? A labor of love by someone who got tired of explaining the same agent architecture patterns over and over? The domain name (agentic-design.ai) is strong, the catalog scale is ambitious, and the freemium model is coherent. But there's no About page, no team page, no blog. The absence of author identity is conspicuous in a field where the best resources are personally signed.
Sources: raw/agentic-design Last updated: 2026-08-01