ShinkaEvolve

LLMs as mutation operators for evolutionary code optimization. Sakana AI, 1.1k stars, Apache-2.0. Maintains populations of programs that evolve toward a fitness function.

I used this heavily ~6 months ago. Revisiting: major API refactor (Mar 2026), PyPI package, agent skills, ICLR 2026 acceptance.

Source: github.com/SakanaAI/ShinkaEvolve

How it works

Multi-island evolutionary algorithm where LLMs propose code changes:

1. Population – archive of programs (default 40), split across islands
2. Selection – elite parents chosen (top 30%), weighted/power-law/beam search
3. Mutation – LLM generates patches: diff (60%), full rewrite (30%), crossover (10%)
4. Evaluation – program runs run_experiment(), returns combined_score
5. Migration – best programs move between islands every N generations

Three patch types keep diversity: small diffs for local search, full rewrites for exploration, crossover for combining good ideas from different lineages.

What changed since Oct 2025

Key design choices

Model ensemble with cost-aware UCB bandit: rotates between gpt-5-mini, gemini-3-flash, gemini-3.1-pro, gpt-5.4. Cheaper models get more tries, expensive ones only when they prove better. Budget cap in USD.

Meta-evolution: the system prompt itself evolves every N generations. Archive of 10 prompts, UCB sampling. The prompts that produce better mutations survive.

Novelty search: code embedding similarity (text-embedding-3-small) prevents the population from collapsing to one solution. Threshold 0.99, optional LLM judgment.

Async 5-10x speedup: proposals and evaluations run concurrently with EWMA smoothing for adaptive pacing.

How we use it

Our skills (/shinka-setup, /shinka-run, /shinka-inspect, /shinka-convert) wrap ShinkaEvolve for agent-bit prompt evolution and code optimization. The prompts/system.md in agent-bit was evolved through ShinkaEvolve runs.

This is the production implementation of self-evolution: the fixed/mutable partition. System prompt is mutable, tool definitions are fixed. ShinkaEvolve searches the mutable space.

Connects to erc3-agents-competition: ERC3 used evolution (68→100%) on agent configurations. ShinkaEvolve is the general-purpose version of that pattern.

And to asi-evolve-ai-research-agents: both use LLMs in evolutionary loops, but ASI-Evolve targets model architectures while ShinkaEvolve targets code/prompts.

Quick start

pip install shinka-evolve
shinka_launch variant=circle_packing_example

Or via our skills:

/shinka-setup    # scaffold evaluate.py + initial.py
/shinka-run      # launch evolution
/shinka-inspect  # load top programs into context

Links

• GitHub – 1.1k stars, Apache-2.0
• Docs
• ICLR 2026 paper
• Agent skills – npx skills add SakanaAI/ShinkaEvolve