darwinian-evolver – skill Dữ liệu & Khoa học cho agent AI

Darwinian Evolver

Run Imbue's [darwinian_evolver](https://github.com/imbue-ai/darwinian_evolver) — an

LLM-driven evolutionary search loop — to optimize a **prompt, regex, SQL query,

or small code snippet** against a fitness function.

Status: thin wrapper around the upstream tool. The skill installs it, walks the

agent through writing a Problem definition (organism + evaluator + mutator),

and drives the loop via the upstream CLI or a small custom Python driver.

License: the upstream tool is AGPL-3.0. The skill ONLY ever invokes it

via the upstream CLI or a subprocess/uv run call (mere aggregation). Do NOT

import upstream classes into Hermes itself.

When to Use

• User says "optimize this prompt", "evolve a regex for X", "auto-improve this

code/SQL", "search for a better instruction".

• You have a scorer (exact match, regex pass-rate, unit test, LLM-judge, runtime

metric) AND a starting candidate (organism). If you don't have a scorer, stop

and define one first — that's the hard part.

• Cost is OK: a typical run is 50–500 LLM calls. On gpt-4o-mini that's pennies;

on Claude Sonnet it can be a few dollars.

Do not use this when:

• The optimization target is differentiable (use gradient descent / DSPy).
• You only need to try 2–3 variants — just write them by hand.
• The fitness signal is purely subjective with no measurable criterion.

Prerequisites

• Python ≥3.11
• git, uv (or pip)
• One of: OPENROUTER_API_KEY, ANTHROPIC_API_KEY, or OPENAI_API_KEY

The skill ships a small parrot_openrouter.py driver that uses OPENROUTER_API_KEY

via the OpenAI SDK, so any model on OpenRouter works. The upstream CLI itself

hardcodes Anthropic and needs ANTHROPIC_API_KEY.

Install (One-Time)

Run via the terminal tool:

mkdir -p ~/.hermes/cache/darwinian-evolver && cd ~/.hermes/cache/darwinian-evolver
[ -d darwinian_evolver ] || git clone --depth 1 https://github.com/imbue-ai/darwinian_evolver.git
cd darwinian_evolver && uv sync

Verify:

cd ~/.hermes/cache/darwinian-evolver/darwinian_evolver \
  && uv run darwinian_evolver --help | head -5

Quick Start — The Built-In Parrot Example

Tiny smoke test (requires ANTHROPIC_API_KEY):

cd ~/.hermes/cache/darwinian-evolver/darwinian_evolver
uv run darwinian_evolver parrot \
  --num_iterations 2 \
  --num_parents_per_iteration 2 \
  --mutator_concurrency 2 --evaluator_concurrency 2 \
  --output_dir ~/.hermes/cache/scratch/parrot_demo

Outputs:

• ~/.hermes/cache/scratch/parrot_demo/snapshots/iteration_N.pkl — pickled population per iteration
• ~/.hermes/cache/scratch/parrot_demo/<jsonl> — per-iteration JSON log (path printed at end)

Open ~/.hermes/cache/darwinian-evolver/darwinian_evolver/darwinian_evolver/lineage_visualizer.html

in a browser and load the JSON log to see the evolutionary tree.

Quick Start — OpenRouter Driver (No Anthropic Key)

The skill ships scripts/parrot_openrouter.py — same parrot problem, but the

LLM call goes through OpenRouter so any provider works.

# From wherever the skill is installed:
SKILL_DIR=~/.hermes/skills/research/darwinian-evolver
DE_DIR=~/.hermes/cache/darwinian-evolver/darwinian_evolver

cd "$DE_DIR" && \
  EVOLVER_MODEL='openai/gpt-4o-mini' \
  uv run --with openai python "$SKILL_DIR/scripts/parrot_openrouter.py" \
    --num_iterations 3 --num_parents_per_iteration 2 \
    --output_dir ~/.hermes/cache/scratch/parrot_or

Inspect the result with scripts/show_snapshot.py:

uv run --with openai python "$SKILL_DIR/scripts/show_snapshot.py" \
  ~/.hermes/cache/scratch/parrot_or/snapshots/iteration_3.pkl

Expected output: 7 evolved prompt templates ranked by score, with the best

landing around 0.6–0.8 (the seed Say {{ phrase }} scored 0.000).

Defining a Custom Problem

The skill ships templates/custom_problem_template.py — copy, edit, run.

Three things you must define:

1. Organism — a Pydantic BaseModel subclass holding the artifact being

evolved (prompt_template: str, regex_pattern: str, sql_query: str,

code_block: str, etc.). Add a run(*args) method that exercises it.

2. Evaluator — .evaluate(organism) -> EvaluationResult(score=..., trainable_failure_cases=[...], holdout_failure_cases=[...], is_viable=True).

• score is in [0, 1]. Higher is better.
• trainable_failure_cases — what the mutator sees. Include enough

context (input, expected, actual) for the LLM to diagnose.

• holdout_failure_cases — kept out of the mutator's view. Use these

to detect overfitting.

• is_viable=True unless the organism is completely broken (raises,

returns None, etc.). A 0-score viable organism is fine — it just gets

down-weighted in parent selection.

3. Mutator — .mutate(organism, failure_cases, learning_log_entries) -> list[Organism].

Typically: build an LLM prompt that includes the current organism + a

failure case + an ask to propose a fix; parse the LLM's response; return

a new Organism. Return [] on parse failure — the loop handles it.

Then write a driver script that wires Problem(initial_organism, evaluator, [mutators])

into EvolveProblemLoop and iterates over loop.run(num_iterations=N) — the

shipped scripts/parrot_openrouter.py is the reference.

Hyperparameters That Actually Matter

| flag | default | when to change |

|---|---|---|

| --num_iterations | 5 | bump to 10–20 once you trust the evaluator |

| --num_parents_per_iteration | 4 | drop to 2 for cheap exploration |

| --mutator_concurrency | 10 | drop to 2–4 to avoid rate limits |

| --evaluator_concurrency | 10 | same; evaluator hits the LLM too |

| --batch_size | 1 | raise to 3–5 once your mutator handles multiple failures |

| --verify_mutations | off | turn on once mutator is wasteful (>10× cost saving on later runs per Imbue) |

| --midpoint_score | p75 | leave alone unless scores cluster |

| --sharpness | 10 | leave alone |

Pitfalls

1. Initial organism must be viable — set is_viable=True in your

EvaluationResult even on a 0-score seed. The loop refuses non-viable

organisms because they imply the loop has nothing to evolve from.

2. Provider content filters kill runs. Azure-backed OpenRouter models

reject phrases like "ignore previous instructions" with HTTP 400. Wrap

the LLM call in try/except and return f"<LLM_ERROR: {e}>" — the

evolver will just score that organism 0 and move on.

3. loop.run() is a generator — calling it doesn't run anything until

you iterate. Use for snap in loop.run(num_iterations=N):.

4. Snapshots are nested pickles. iteration_N.pkl contains a dict with

population_snapshot (more pickled bytes). To unpickle you must have the

Organism class importable under the same dotted path it was pickled at.

5. Concurrency defaults are aggressive. 10/10 will hit rate limits on

most providers. Start with 2/2.

6. CLI is hardcoded to Anthropic. uv run darwinian_evolver <problem>

reaches for ANTHROPIC_API_KEY and uses Claude Sonnet. To use any other

provider, write a driver like parrot_openrouter.py.

7. AGPL. Never from darwinian_evolver import ... inside Hermes core.

Custom driver scripts under ~/.hermes/skills/... are user-side and fine.

8. No PyPI package. pip install darwinian-evolver will pull the wrong

thing. Always install from the GitHub repo.

Verification

After install + a parrot run, exit code 0 from this is sufficient:

DE_DIR=~/.hermes/cache/darwinian-evolver/darwinian_evolver
ls "$DE_DIR/darwinian_evolver/lineage_visualizer.html" >/dev/null && \
cd "$DE_DIR" && uv run darwinian_evolver --help >/dev/null && \
echo "darwinian-evolver: OK"

References

• [Imbue research post](https://imbue.com/research/2026-02-27-darwinian-evolver/)
• [ARC-AGI-2 results](https://imbue.com/research/2026-02-27-arc-agi-2-evolution/)
• [imbue-ai/darwinian_evolver](https://github.com/imbue-ai/darwinian_evolver) (AGPL-3.0)
• [Darwin Gödel Machines](https://arxiv.org/abs/2505.22954)
• [PromptBreeder](https://arxiv.org/abs/2309.16797)