Skip to content
Live specimen · evolution chamber

Code that breeds better code

This page is not a demo. A real evolutionary engine is running in your browser right now — mutations competing across island populations, MAP-Elites preserving diversity, the fittest surviving. That's the product. Define what "better" means; evolution does the rest. Zero prompts in the loop.

8 gen/s
specimen · damped wave algorithm_discovery
0
Generation
0.000
Best Fitness
0
Evaluations
0%
Grid Coverage
0
Migrations
STATION 01Instruments — all readings live

Every gauge below is wired
to the same running engine.

Watch a session unfold. Mutations compete. Fitness climbs. The grid fills. All autonomous — and all real: these instruments read from the evolution running in this page, not a recording.

event streamws://session
best program · livegen 0
fitness / precisionper generation
map-elites archiveclick a cell to inspect
← simplercomplexity →
STATION 02The shift
Traditional AI coding
1. Write prompt
2. Get one answer
3. Review manually
4. Tweak prompt
5. Repeat forever
Human is the loop
EvolveHandsAI
1. Define fitness function
2. Launch evolution
3. Hundreds of LLM mutations
4. Survival of the fittest
5. Get the best code
Evolution is the loop
STATION 03Capabilities

Not a code assistant.
An optimization engine.

Searches the space of possible implementations. Maintains diverse solutions. Rejects regressions. Runs unsupervised.

SPEC-01

LLM Ensemble

Claude, GPT-4o, Haiku generate competing mutations. Adaptive weighting shifts traffic to models producing better variants.

SPEC-02

MAP-Elites Grid

Quality-diversity algorithm maintains a grid of solutions indexed by feature dimensions. Finds approaches you'd never think to try.

▶ live: archive 0% full in this page's run
SPEC-03

Island Parallelism

2–5 isolated populations evolve independently. Periodic migration shares breakthroughs without homogenizing the search.

▶ live: 0 migrations so far in this page's run
SPEC-04

Regression Detection

EvoClaw-inspired precision tracking. Candidates that break existing tests are rejected. F1 balances recall and precision.

▶ live: offspring precision 100%
SPEC-05

Milestone Evolution

Dependency-graph milestones for long-horizon tasks. Prior milestones become regression tests automatically.

SPEC-06

Docker Sandbox

Read-only filesystem, no network, 256MB RAM, 1 CPU. Safely evolve untrusted code in full isolation.

STATION 04Protocol

Five steps. Then sleep.

Mark the code, define fitness, launch. Evolution runs autonomously.

01

Mark the mutable region

Wrap code with EVOLVE-BLOCK markers. Everything outside stays frozen.

02

Write a fitness function

A script that takes a program and outputs a score. Accuracy, speed, compression — anything measurable.

03

Pick a preset & launch

quick_prototype for 30 generations in a minute. evoclaw_evolution for 300 overnight.

04

Watch evolution unfold

Real-time WebSocket streaming. Fitness curves. MAP-Elites grid filling. Island migrations. Or just wait.

05

Retrieve the winner

Get the best program with full lineage, feature coordinates, and evaluation metrics. It won't be what you'd write.

pipeline.py
def pipeline(data):
    cleaned = preprocess(data)

    # EVOLVE-BLOCK-START
    def optimize(x):
        return sorted(x, key=lambda r: r.score)
    # EVOLVE-BLOCK-END

    return postprocess(optimize(cleaned))
launch.sh
curl -X POST localhost:3000/api/v1/evolve/start \
  -H "Content-Type: application/json" \
  -d '{"preset": "algorithm_discovery",
       "session_id": "my-run"}'
STATION 05Presets — load one into the chamber

Pick a strategy.
Launch in seconds.

Each preset configures model weights, temperatures, population sizes, cascade stages, and evaluation timeouts. Press "run in chamber" to load its search profile into the live engine above.

quick_prototype
Fast iteration for testing. Haiku-heavy ensemble, small populations.
30 gens2 islands15s
algorithm_discovery
Broad exploration for novel algorithms. High temperature, large populations.
200 gens5 islands60s
ml_research
Fixed 5-minute evaluation budget for ML training optimization.
100 gens3 islands5m
performance_optimization
Heavy exploitation, low temperature. Targeted speed and efficiency.
150 gens3 islands30s
evoclaw_evolutionNEW
Long-horizon continuous evolution with regression detection. Docker sandbox. Precision as a feature dimension.
300 gens4 islandsDocker
STATION 06Infrastructure

Production-grade.
From day one.

security layers
  • Security scanner blocks eval(), os.system(), subprocess
  • Docker sandbox: read-only FS, no network, 256MB RAM
  • Per-session token and cost budgets with 80% warnings
  • Rate limiting: 5/user, 50 global (Redis-backed)
  • Input validation on all config values and code sizes
observability
  • Full audit trail: every mutation, evaluation, migration
  • WebSocket streaming with auto-reconnect + polling fallback
  • /health endpoint for Kubernetes liveness probes
  • /metrics endpoint with per-session aggregate stats
  • Session persistence and restore after restart
0
API Endpoints
0
Tests Passing
0
Presets
0
Modules

Stop writing code.
Start evolving it.

Let hundreds of LLM-generated variants compete and breed across generations. Get solutions no single prompt could produce.

Get Started Star on GitHub
GEN
0
BEST
0.000