Launched this week
Agentic Design
288 AI agent design patterns, with code that actually runs
2 followers
288 AI agent design patterns, with code that actually runs
2 followers
A free catalog of 288 agentic AI design patterns: TLDRs, diagrams, interactive demos, runnable TypeScript/Python/Rust code, verified references. A $29 pack turns it into an MCP server for Claude Code and Cursor. Pro adds an AI expert, lessons and an eval lab.
Hi Product Hunt! I'm David, a software engineer based in France. Agentic Design is my attempt to fix how we learn to build AI agents.
Most tutorials teach a framework. The framework changes, and what you learned goes with it. So I built the opposite: a catalog of the underlying patterns, each implemented from scratch in TypeScript, Python and Rust. No frameworks, no glue code. You read the mechanism, run it on the page, and keep the idea when the tooling churns.
What's free, no login:
- 288 patterns across 24 categories: reasoning, routing, memory, multi-agent coordination, evaluation, agent UX, and interop protocols like MCP, A2A and x402
- Every pattern has a 30-second TLDR, a flow diagram, an interactive demo built for that specific pattern, readable code in all three languages, and the tradeoffs that never make it into a README
- Also free: a red-teaming hub with 108 attack techniques against agents, 22 model architecture families with a selection matrix, a 20-chapter guide to the inference stack, and a daily narrated AI news briefing
- The whole catalog in 6 languages, plus an llms-full.txt if you'd rather hand it to a model
Running the code needs a free account: 10 runs a day, each in a throwaway isolated sandbox, about two seconds a run. Every sample has been executed end to end, and every reference links to a real paper.
Where the line is:
- Patterns Pack, $29 one time: the catalog as an MCP server plus skills, rules and slash commands for Claude Code and Cursor, so your coding agent consults all 288 patterns while it writes
- Pro, $29/mo: an AI expert grounded on the catalog, an eval lab that runs one task across patterns and models with real latency and cost numbers, and 92 guided lessons with challenges that execute in the same sandbox
Everything that's pure explanation is free. Everything compute-backed is paid.
For launch week: LAUNCH50 gets you 50% off your first payment, Pro or the Pack, until August 20.
I would honestly love to hear where the patterns are wrong, and which ones are missing. Which pattern do you find yourself reaching for most?