Living Pattern Catalogs That Keep Your Architecture Honest

Simon Willison maintains a living guide to agentic engineering patterns — not a blog post that decays, but a document he updates as the field moves. That distinction matters. Most writing about agent architecture is a snapshot. This is a rolling audit of what actually works, maintained by someone who builds with these tools daily and writes about them with uncommon rigor.

The value here isn’t just “smart person wrote a patterns doc.” It’s that a well-maintained, opinionated catalog of patterns from a credible practitioner becomes prior art you can pressure-test your own decisions against. Every architectural choice in joelclaw’s agent loop system — the five separated roles, the Inngest step-based durability, the fan-out and cross-checking — either aligns with a documented pattern or deliberately diverges from one. Both are useful signals. Alignment means you’re not inventing where you don’t need to. Divergence means you’d better have a reason.

Simon’s broader body of work on LLM tooling, Datasette, and the shot-scraper ecosystem reflects the same philosophy: build small composable tools, document them publicly, iterate in the open. That’s the same bet behind joelclaw’s CLI-first, event-driven approach. The patterns guide is worth monitoring not because it tells you what to build, but because it tells you what other people are converging on — and where you might be building something nobody else needs.

Worth revisiting quarterly as a sanity check against your own ADRs.

Key Ideas

• Living documents beat blog posts for fast-moving domains — a guide that updates is more useful than a post that was correct six months ago
• Pattern catalogs as architecture mirrors — comparing your system’s choices against a canonical list surfaces both validation and blind spots
• Simon Willison’s documentation practice is itself a pattern worth studying — public, iterative, tool-assisted, high-volume, and relentlessly honest
• Agentic engineering is converging on shared vocabulary — role separation, tool use, human-in-the-loop gates, multi-model verification — and this guide tracks that convergence
• Composable small tools over monolithic agents is a recurring theme in Simon’s work, echoing joelclaw’s CLI-first and event-bus architecture

Links

• Agentic Engineering Patterns Guide — the source, living document
• Simon Willison’s Blog — prolific writing on LLMs, tools, and open source
• LLM CLI Tool — Simon’s command-line tool for working with language models
• Datasette — Simon’s tool for exploring and publishing data
• shot-scraper — browser automation tool, same composable-tools philosophy
• Simon Willison on GitHub
• Simon Willison on Mastodon