Skip to main content
Open Source

Code that improves itself

We don't just build websites. We build the tools that make AI systems better. Open source. TypeScript. Runs in our own agent fleet.

Flagship Project

Darwin Agents

A TypeScript framework for AI agents that improve their own prompts, automatically, without manual tuning. A critic scores every run. Better variants are A/B-tested with always-valid statistics. Anything that regresses rolls back. The key difference from other optimization approaches: the reflective loop runs continuously while agents work, not as a periodic batch job.

TypeScriptMIT LicenseMulti-ProviderSQLite + PostgreSQL

0

Hard deps

456

Tests

MIT

License

terminal
# Install
npm install darwin-agents

# Run your first self-evolving agent
npx darwin run writer \
  "Explain quantum computing simply"

# Watch it evolve
npx darwin evolve writer
→ Pattern detected: market topics below target
→ New variant: writer-v2 generated
→ A/B test: v2 wins (7.12 vs 6.89 / 10)
✓ writer-v2 is now default

Self-Evolution

Agents detect weakness patterns and automatically generate improved prompt variants.

A/B Testing

New variants are tested against the current version. The winner becomes the default.

Safety Gates

Every evolution passes through quality gates. No uncontrolled prompt drift.

Multi-Provider

Use any model. Claude, OpenAI, Ollama, or your own. Darwin doesn't lock you into a provider. Swap the model, keep the evolution loop.

How it works

The Evolution Loop

Darwin improves AI agents automatically. No manual prompt engineering required.

1

Agent runs a task

Your agent handles an assignment, writing, researching, analyzing.

2

Critics evaluate quality

Specialized critic agents measure quality, sources, structure, and relevance.

3

Patterns emerge

After enough runs, Darwin detects patterns: 'weak on technical topics', 'too verbose in summaries'.

4

New variant is generated

An optimizer generates an improved prompt variant based on detected patterns.

5

A/B test decides

The new variant is tested against the current version. Blind, fair, data-driven.

6

Winner becomes default

The better variant takes over. Your agent got better, automatically.

Get involved

Built in the open, because that's how you earn trust

Darwin is MIT-licensed. We publish the source, the tests, and the results. Issues and pull requests are welcome.