Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Get Started Free →

evolve — Darwin Mode self-improvement · ruvnet | DecimalAI

agent toolingby ruvnet

evolve — Darwin Mode self-improvement

ruvnet/evolve

Source ruvnet/metaharness ↗License MIT

Evolve this harness with Darwin Mode — frozen model, evolving harness (real, sandboxed, safety-gated).

SkillScore96Maintained

Automatic

AI Engineer pack

Last verified 9/1/2026

What it changes

Passed 23 of 24 test cases with the skill, and 4 of 24 without it.

Without skill16%

With skill96%

Checks it passed that the model alone missed

Recommends using npm run evolve:dry.
Mentions guarding against Goodharting.
Cites approximately doubling the resolve rate.

and 30 more in the full report

20 fixed4 passed either way1 still failing0 broke

Pass-rate lift: +80 pp
Tokens: -13%
Agent turns: 0%
Response time: -42%

Model gemini-3.6-flashMeasured 9/1/2026v124 of 25 cases passed with skill · 24 comparable

Tokens, turns and time describe resource use, not a dollar price.

Use this skill

Save it into your project's skills folder. Claude Code picks it up in its next session.

curl -fsSL --create-dirs -o .claude/skills/ruvnet-evolve/SKILL.md https://app.decimal.ai/s/ruvnet-evolve/SKILL.md

Want it pinned and updatable? pip install "decimalai>=0.15", then decimalai skills add ruvnet-evolve, and start Claude Code with claude --plugin-dir "$(git rev-parse --show-toplevel)/.agents"

1 · One-time — install the CLI (docs ↗)

pip install "decimalai>=0.15"

2 · Get the skill — writes .agents/skills/ruvnet-evolve/ and pins it in .decimal/skills.lock

decimalai skills add ruvnet-evolve

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/ruvnet-evolve/SKILL.md https://app.decimal.ai/s/ruvnet-evolve/SKILL.md

Raw file: /s/ruvnet-evolve/SKILL.md ↗ · machine-readable summary at /s/ruvnet-evolve.json ↗

Safety checks: 3 of 3 passed

Code scan CleanIntent PassedContent Clean

See it in action

One of 19 test cases the skill fixed. Same model, same prompt, with and without the skill.

1 of 19

Prompt

I want to invoke the Darwin harness evolution CLI directly to run 3 generations with 4 child variants per generation in a real sandbox environment targeting the current directory. What exact CLI command line invocation executes this?

What was checked, without and with the skill
Checked	Without skill	With skill
Includes npx metaharness-darwin evolve in the CLI command.	Missed	Passed
Includes --generations 3 in the CLI invocation.	Passed	Passed
Includes --children 4 in the CLI invocation.	Passed	Passed

case-07 · gemini-3.6-flash · 9/1/2026 · answers are capped and redacted

What's inside

The instructions your model reads when the skill loads.

SKILL.md48 linessingle file

evolve — Darwin Mode self-improvement

kimi-k3-harness ships with Darwin Mode (@metaharness/darwin, ADR-070…146): the model is frozen; the harness evolves. Each generation mutates ONE of the 7 surface files (planner, contextBuilder, reviewer, retry/tool/memory/score policy), sandboxes each child, scores it, and keeps only variants that measurably improve — building an archive of successful descendants.

Run it

npm run evolve        # real substrate: runs your test command per variant (deterministic mutator — no API key, no network)
npm run evolve:dry    # mock substrate: fast, fully offline, no test execution

Or directly:

npx metaharness-darwin evolve . --sandbox real --generations 3 --children 4

Safety (secure by default)

Deterministic mutator is the default — no network, no API key, air-gapped.
Every mutation passes the validateGeneratedCode gate: no new imports, network, filesystem, shell, env access, or dependencies — pure refactor/tuning only.
Mutations run in a sandbox; only variants that pass your tests are archived.
Nothing is promoted without measured improvement (guard against Goodharting).

See @metaharness/darwin for selection strategies (--selection, --crossover, --curriculum), statistical gates (--fdr, --bench), and the real-LLM mutator (library API).

What the benchmarks taught us (measured, full SWE-bench Lite 300)

Defaults worth carrying into how you evolve and run this harness (full evidence + CIs in @metaharness/darwin's LEARNINGS.md / bench/results/RESULTS.md):

Closed-loop repair is the #1 lever (~2×). Feeding test/compiler failure back and retrying took resolve-rate 7.7% → 15.3% on the same cheap model. Iterate against ground truth, don't single-shot.
Cheap-first + cost-aware routing. Track $/resolve, not just resolve-rate; a cheap model

Track record

Benchmark

23 / 24

test cases passed, vs 4 / 24 without

Adoption

Users: 0
Times used: 0

Safety and source details

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/10/2026

View 15 scanner checks

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

ruvnet / metaharness

kimi-k3-harness/.claude/skills/evolve/SKILL.md

View on GitHub ↗

License MITImported 8/10/2026Version v1

More skills from the same source and nearby in the registry, with their headline benchmark lift.

Ready to supercharge your AI agents?

See it in action

What's inside

evolve — Darwin Mode self-improvement

Run it

Safety (secure by default)

What the benchmarks taught us (measured, full SWE-bench Lite 300)

Track record

Benchmark

Adoption

SkillSafety

Source

Related skills