Agent Self-Evolution — Automated Evaluation & Improvement

⭐ If you find this useful, please star the GitHub repo before installing — it helps others discover it.

Your agent does a lot of things. But is it getting better? Or are your latest tweaks breaking things you didn't notice?

Agent Self-Evolution gives you a framework to measure, test, and systematically improve your agent — like unit tests, but for agent behavior.

What's inside

Golden Test Sets: Define scenarios your agent must handle correctly. Run them periodically and catch regressions before users do.

Ablation Testing: Wondering if that 200-line system prompt section actually helps? Remove it, measure the impact, put it back. Now you know. We found that 7% of one config file was load-bearing for the entire system — without ablation, you'd never know which 7%.

Multi-Dimensional Evaluation: Don't just check pass/fail. Score across dimensions — safety compliance, tool routing accuracy, output quality, memory utilization. Track trends over weeks.

Automated Improvement Loops: Evaluation → identify weakest dimension → targeted fix → re-evaluate. Like gradient descent for agent behavior.

Install

bash {baseDir}/scripts/install.sh

Quick start

from agent_evolution.golden_test import GoldenTestRunner
from agent_evolution.ablation import AblationExperiment

# Define a golden test
runner = GoldenTestRunner()
runner.add_case(
    name="handles-ambiguous-request",
    input="do the thing",
    expected_behavior="asks for clarification rather than guessing",
    dimensions=["safety", "output_quality"]
)

# Run and score
results = runner.run(model="your-agent-endpoint")
print(results.summary())  # Pass rate, dimension scores, regressions

# Ablation: what happens without memory files?
experiment = AblationExperiment(
    baseline_config="agent.yaml",
    conditions={"no_memory": {"remove": ["memory/*.md"]}},
    test_set=runner.cases
)
experiment.run()  # Measures impact of each ablation

Key findings from our own agent

• SOUL.md (7% of config by characters): removing it caused system-wide behavioral collapse (Cohen's d = 0.602) — it's not fluff, it's load-bearing
• Memory files: most essential component (d = 0.944) — without history, the agent becomes generic
• Safety rules: removal didn't just reduce safety — it degraded all dimensions (d = 0.609)

Companion projects

• nous-safety — Runtime safety engine with Datalog reasoning
• biomorphic-memory — Brain-inspired memory with spreading activation

Requirements

• Python ≥ 3.11
• An LLM API key for evaluation judging (strong model recommended — GPT-5.4 / Opus)

License

Apache 2.0