Evolving AI Agents with A-Evolve

Overview

A-Evolve is universal infrastructure for evolving any AI agent across any domain using any evolution algorithm with zero manual engineering. It represents all evolvable agent state as files (prompts, skills, memory, tools), runs iterative solve-observe-evolve cycles against benchmarks, and uses LLM-driven mutation to improve agent performance automatically.

Benchmark results (Claude Opus 4.6):

• MCP-Atlas: 79.4% (#1)
• SWE-bench Verified: 76.8% (~#5)
• Terminal-Bench 2.0: 76.5% (~#7)
• SkillsBench: 34.9% (#2)

When to Use A-Evolve

Use A-Evolve when:

• Optimizing agent prompts, skills, or memory against a measurable benchmark
• Building self-improving agents with automated gating and rollback
• Evolving domain-specific tool usage and procedures through LLM-driven mutation
• Running iterative solve-observe-evolve loops to maximize agent performance
• Needing reproducible, git-versioned evolution history for every change

Key differentiator: Other frameworks _build_ agents; A-Evolve _optimizes_ them. It sits on top of any agent framework and makes it better through automated evolution.

Do NOT use A-Evolve for:

• Building multi-agent orchestration from scratch (use CrewAI, LangGraph)
• One-shot agent tasks with no iteration needed (use LangChain, LlamaIndex)
• RAG pipeline optimization (use LlamaIndex, Chroma)
• Prompt-only optimization without skill/memory evolution (use DSPy)

Quick Start

Installation

bash
pip install a-evolve                    # Core
pip install a-evolve[anthropic]         # With Claude support
pip install a-evolve[all]               # All providers

Three-Line Evolution

python
import agent_evolve as ae

evolver = ae.Evolver(agent="swe", benchmark="swe-verified")
results = evolver.run(cycles=10)
print(f"Final score: {results.final_score}")

This copies the built-in SWE seed workspace, runs 10 evolution cycles against SWE-bench Verified, and returns the optimized agent.

Core Concepts

The Agent Workspace

All evolvable state lives as files in a workspace directory:

my-agent/
├── manifest.yaml          # Metadata + entrypoint
├── prompts/
│   ├── system.md          # Main system prompt (evolved)
│   └── fragments/         # Modular prompt pieces
├── skills/
│   └── skill-name/
│       └── SKILL.md       # Reusable procedure with frontmatter
├── memory/
│   ├── episodic.jsonl     # Lessons from failures
│   └── semantic.jsonl     # General knowledge
├── tools/
│   ├── registry.yaml      # Tool manifest
│   └── tool_name.py       # Tool implementations
└── evolution/             # Managed by engine (metrics, history)

The Evolution Loop

Each cycle follows five phases:

1. Solve — Agent processes a batch of tasks from the benchmark
2. Observe — Benchmark evaluates trajectories, producing (task, trajectory, feedback) triples
3. Evolve — Evolution engine mutates workspace files based on observations
4. Gate — Validate mutations (git snapshot before/after for rollback)
5. Reload — Agent reinitializes from evolved filesystem state

Three Pluggable Interfaces

python
# 1. Agent — implements solve()
class MyAgent(ae.BaseAgent):
    def solve(self, task: ae.Task) -> ae.Trajectory:
        # Domain-specific solving logic
        return ae.Trajectory(task_id=task.id, output=result, steps=steps)

# 2. Benchmark — implements get_tasks() and evaluate()
class MyBenchmark(ae.BenchmarkAdapter):
    def get_tasks(self, split="train", limit=None) -> list[ae.Task]:
        return [ae.Task(id="1", input="...")]

    def evaluate(self, task: ae.Task, trajectory: ae.Trajectory) -> ae.Feedback:
        return ae.Feedback(success=True, score=0.95, detail="Passed")

# 3. Engine — implements step()
class MyEngine(ae.EvolutionEngine):
    def step(self, workspace, observations, history, trial):
        # Mutate workspace based on observations
        return ae.StepResult(mutated=True, summary="Updated prompts")

Workflow 1: Evolve an Existing Agent

Use when: You have a working agent and want to optimize it against a benchmark.

Critical Requirements:

• ] Agent implements BaseAgent.solve() returning Trajectory
• ] Benchmark implements BenchmarkAdapter with get_tasks() and evaluate()
• ] Seed workspace has manifest.yaml with entrypoint and evolvable layers
• ] System prompt exists at prompts/system.md
• ] Workspace is a git repo (run git init && git add -A && git commit -m "init")

Steps

python
import agent_evolve as ae

# Configure evolution parameters
config = ae.EvolveConfig(
    batch_size=10,           # Tasks per solve round
    max_cycles=20,           # Maximum evolution iterations
    evolve_prompts=True,     # Mutate system prompt
    evolve_skills=True,      # Discover and refine skills
    evolve_memory=True,      # Build episodic memory
    evolver_model="us.anthropic.claude-opus-4-6-v1",
)

# Point to your agent workspace and benchmark
evolver = ae.Evolver(
    agent="./my-agent-workspace",
    benchmark="swe-verified",     # Or custom BenchmarkAdapter instance
    config=config,
)

# Run evolution
results = evolver.run(cycles=10)

# Inspect results
print(f"Cycles completed: {results.cycles_completed}")
print(f"Final score: {results.final_score}")
print(f"Converged: {results.converged}")
for cycle_num, score in enumerate(results.score_history):
    print(f"  Cycle {cycle_num + 1}: {score:.3f}")

Post-Evolution

The workspace is now optimized. Inspect what changed:

bash
cd my-agent-workspace
git log --oneline              # See evo-1, evo-2, ... tags
git diff evo-1 evo-10          # Compare first and last evolution
cat prompts/system.md          # Read evolved prompt
ls skills/                     # See discovered skills

Workflow 2: Add a Custom Benchmark

Use when: You want to evolve agents on your own domain-specific tasks.

Critical Requirements:

• ] Define task format (inputs, expected outputs)
• ] Implement scoring logic (0.0–1.0 scale)
• ] Prepare task dataset (train + holdout split)

Steps

python
import agent_evolve as ae

class CodeReviewBenchmark(ae.BenchmarkAdapter):
    """Evaluate agents on code review quality."""

    def get_tasks(self, split="train", limit=None):
        tasks = load_review_dataset(split)
        if limit:
            tasks = tasks[:limit]
        return [
            ae.Task(id=t["id"], input=t["diff"], metadata={"expected": t["comments"]})
            for t in tasks
        ]

    def evaluate(self, task, trajectory):
        expected = task.metadata["expected"]
        actual = trajectory.output
        precision, recall = compute_review_metrics(expected, actual)
        f1 = 2 * precision * recall / (precision + recall + 1e-9)
        return ae.Feedback(
            success=f1 > 0.7,
            score=f1,
            detail=f"P={precision:.2f} R={recall:.2f} F1={f1:.2f}",
        )

# Use with any agent
evolver = ae.Evolver(agent="./my-agent", benchmark=CodeReviewBenchmark())
results = evolver.run(cycles=5)

Workflow 3: Create a Custom Evolution Engine

Use when: The default LLM-driven mutation doesn't suit your domain.

Steps

python
import agent_evolve as ae

class RuleBasedEngine(ae.EvolutionEngine):
    def step(self, workspace, observations, history, trial):
        failures = [o for o in observations if not o.feedback.success]
        if not failures:
            return ae.StepResult(mutated=False, summary="No failures to address")

        # Analyze failure patterns
        error_types = categorize_errors(failures)
        prompt = workspace.read_prompt()

        # Append learned rules to prompt
        new_rules = generate_rules(error_types)
        workspace.write_prompt(prompt + "\n" + new_rules)

        return ae.StepResult(
            mutated=True,
            summary=f"Added {len(new_rules)} rules from {len(failures)} failures",
        )

evolver = ae.Evolver(
    agent="./my-agent",
    benchmark="my-benchmark",
    engine=RuleBasedEngine(),
)

Built-in Components

Seed Agents

| Agent | Domain | Model | Key Feature | |-------|--------|-------|-------------| | swe | SWE-bench | Claude Opus 4.6 | Verify-fix loop, skill proposals | | terminal | Terminal-Bench | Claude Sonnet 4 | Concurrent timeout, env discovery | | mcp | MCP-Atlas | Claude Opus 4.6 | MCP server integration |

Benchmarks

| Name | Domain | Metric | |------|--------|--------| | swe-verified | Code patching | Pass rate | | mcp-atlas | Tool calling | Accuracy | | terminal2 | Shell tasks | Pass rate | | skill-bench | Multi-step procedures | Accuracy | | arc-agi-3 | Interactive games | RHAE score |

Evolution Algorithms

| Algorithm | Strategy | Best For | |-----------|----------|----------| | A-Evolve/SkillForge | LLM-driven workspace mutation | General-purpose | | Guided Synthesis | Memory-first, curated skills | Skill discovery | | Adaptive Evolution | Reward tracking, filtered observations | Fine-grained control | | Adaptive Skill | Skill-centric refinement | Skill-heavy domains |

Configuration Reference

python
ae.EvolveConfig(
    batch_size=10,              # Tasks per solve round
    max_cycles=20,              # Max evolution iterations
    holdout_ratio=0.2,          # Test set split for gating
    evolve_prompts=True,        # Mutate system prompts
    evolve_skills=True,         # Discover/refine skills
    evolve_memory=True,         # Build episodic memory
    evolve_tools=False,         # Mutate tool implementations
    trajectory_only=False,      # Hide scores from evolver
    evolver_model="us.anthropic.claude-opus-4-6-v1",
    evolver_max_tokens=16384,
    egl_threshold=0.05,         # Convergence epsilon
    egl_window=3,               # Cycles for plateau detection
)

Convergence: Evolution stops early when score improvement is less than egl_threshold over the last egl_window cycles.

Skill Format

Skills are reusable procedures discovered and refined during evolution:

markdown
---
name: verify-edge-cases
description: "TRIGGER when: checking boundary conditions. DO NOT TRIGGER: for happy-path tests."
---

## Pattern
Test all falsy-but-valid values: 0, False, "", [], {}

## Process
1. List all input boundaries
2. Run each against the implementation
3. Check both output AND side effects

Skills accumulate in the workspace skills/ directory. The evolver curates them: ACCEPT new skills, MERGE overlapping ones, SKIP redundant proposals. Target: 5–10 broad skills, not 30 narrow ones.

Common Issues

Evolution score plateaus early

Cause: Batch size too small or evolver doesn't see enough failure diversity. Fix: Increase batch_size (try 15–20) and ensure benchmark tasks cover diverse failure modes. Set trajectory_only=False so the evolver sees scores.

Agent workspace grows too large

Cause: Skill library bloat from accepting every proposal. Fix: The default SkillForge engine curates skills automatically. If using a custom engine, implement merging logic to consolidate overlapping skills.

Git conflicts during evolution

Cause: Multiple evolution runs on the same workspace. Fix: Each evolver.run() should operate on its own workspace copy. Use Evolver(agent="seed-name") to auto-copy the seed each time.

LLM provider errors during evolution

Cause: Rate limits or authentication issues with the evolver model. Fix: Check evolver_model config. For Bedrock, ensure AWS credentials are configured. For Anthropic, set ANTHROPIC_API_KEY.

Custom agent not picking up evolved state

Cause: Agent doesn't implement reload_from_fs(). Fix: Override reload_from_fs() in your BaseAgent subclass to re-read prompts, skills, and memory from the workspace after each evolution cycle.

Usage Instructions for Agents

When this skill is loaded:

1. Read this entire file before implementing any evolution workflow
2. Start with the Quick Start — get a minimal evolution running before customizing
3. Use built-in seeds when possible — "swe", "terminal", "mcp" have battle-tested configurations
4. Always initialize git in custom workspaces before running evolution
5. Check convergence settings — default egl_threshold=0.05 with egl_window=3 may be too aggressive for your domain
6. Inspect evolved state after each run — read prompts/system.md and skills/ to understand what the evolver learned

Pro Tips:

• Set trajectory_only=False (default) so the evolver sees scores — this accelerates learning
• Start with batch_size=10 and adjust based on task diversity
• Use holdout_ratio=0.2 to prevent overfitting to training tasks
• After evolution, git diff evo-1 evo-N shows the cumulative effect of all mutations
• If the evolver isn't finding skills, enrich feedback.detail strings with specific failure reasons

Warning Signs:

• Score oscillating between cycles → benchmark evaluation may be non-deterministic
• Skills directory growing past 15+ skills → engine isn't merging/curating properly
• Prompt growing past 10K chars → evolution is appending without refactoring
• converged=True after 2-3 cycles → increase egl_window and decrease egl_threshold

References

• Architecture deep dive: See references/architecture.md
• API reference: See references/api.md
• Step-by-step tutorials: See references/tutorials.md
• Real-world examples: See references/examples.md
• GitHub issues & solutions: See references/issues.md
• Design patterns: See references/design-patterns.md
• Release history: See references/releases.md