MindSwarm
Open Source · MIT

MindSwarm Learn

An open-source project I built: 12 hands-on challenges that turn Anthropic's courses into a real, working multi-agent product. No toy examples — production patterns from challenge one. It is part of how I keep the craft I bring to clients sharp.

View on GitHub →
The idea

Practice, not theory

Every lesson from Anthropic's courses becomes a real, working feature. You don't read about a pattern — you ship it.

Traditional course

  • Read about prompt engineering
  • Watch a tool-use demo video
  • Study evaluation theory
  • "Now go apply this yourself"

MindSwarm Learn

  • Build a prompt lab with versioning & A/B tests
  • Wire 17 tools across 5 real agents
  • Run code-graded + model-graded evals
  • Every lesson IS the application
The build

Four phases, twelve challenges

Phase 1

Foundation

A Claude client with automatic model fallback, business logic replaced by Claude-powered decisions, token streaming over SSE with image analysis.

Phase 2

Intelligence

A versioned prompt lab with A/B tests, 17 tools wired across 5 agents, an agentic support chatbot, production domain prompts.

Phase 3

Quality

A code-graded evaluation engine, Claude-judges-Claude model evals, and a promptfoo CI pipeline for continuous evaluation.

Phase 4

Mastery

A self-improving prompt optimizer that learns from evaluation history, and a full-stack dashboard tracking progress.

12
challenges
25
API endpoints
17
agent tools
$0
runs without an API key

Free, open-source, MIT.

Clone it, run it, learn from it. It works without an API key — graceful fallbacks everywhere.

Clone on GitHub →