Self-evolving multi-agent framework

Agents that evolve themselves

A MetaAgent orchestrates specialized sub-agents to solve tasks — while generator, evaluator, and optimizer agents continuously grow the tool, skill, and agent ecosystem.

Python 3.12 License MIT Multi-Agent Orchestration Hot-plug Extensions
10+
Component registries
4
Specialist sub-agents
Evolution roles per component
8+
Built-in benchmarks
Why AgentEvolver

Built to orchestrate — and to improve itself

A clean, registry-driven core where every capability is a hot-pluggable component, and dedicated agents grow that ecosystem over time.

🧠

MetaAgent Orchestration

A planning agent decomposes tasks, dispatches subtasks over an Agent Bus, executes in parallel, and evaluates — looping until done.

🧬

Self-Evolution Loop

Generator, evaluator, and optimizer agents create, assess, and refine tools, skills, agents, and environments automatically.

🔌

Hot-plug Extensions

The ExtensionManager layers evolved components on top of built-ins at runtime — versioned, archived, and instantly rollback-able.

🗂️

Registry-driven Core

Tools, agents, prompts, skills, benchmarks, hooks, constraints, memory — each self-registers via a clean mmengine registry.

🔭

Live Trace UI

Watch every agent step through a task in real time in your browser, with full run artifacts written to the work directory.

🏆

Benchmark Suite

Ships with AIME, GPQA, GSM8K, HLE, LeetCode, DeepWeb and more — datasets-first with automatic HuggingFace fallback.

Architecture

One planner, many specialists

The MetaAgent plans and dispatches work over the Agent Bus to specialist sub-agents that run in parallel and return results for evaluation.

Task Input
Text · Image · Audio · Video
MetaAgent · Planning
Plan Dispatch Execute Evaluate
↺ reconcile on conflict — loop until done
Answer
Verified result
◆ Agent Bus ◆
🔎 Deep Researcher
Concurrent search
Search EngineLLM SearchImage SearchAggregation
📊 Deep Analyzer
Comprehensive analysis
Cross-ModelCross-ModalSynthesisVerification
⌨️ Vibe Coding
Code execution
CodingCalculationMCP ToolsVerification
🛠️ Tool Calling
Tool & skill calling
BashSkillsSKILL.mdTodo · Done
The Evolution Engine

A framework that rewrites itself

For every kind of component, three agents form a closed improvement loop — new capabilities are born, judged, and refined without touching the immutable core.

Generator
creates new components
⚖️
Evaluator
assesses quality
🔧
Optimizer
refines the code
Evolving Ecosystem

Immutable core, mutable everything-else

Hand-written built-ins live in src/ and never change. Evolved components live in an external extension/ tree, loaded and versioned by the ExtensionManager — so the system can grow safely and roll back instantly.

🧩 Tools 📘 Skills 🤖 Agents 🌐 Environments
Registry-driven

Every capability is a component

Each module self-registers with an mmengine registry — discoverable, swappable, and evolvable.

@TOOL @AGENT @PROMPT @SKILL @BENCHMARK @DATASET @HOOK @CONSTRAINT @ENVIRONMENT @MEMORY_SYSTEM
Evaluate everything

Benchmarks out of the box

Datasets are read locally first, then snapshot-downloaded from HuggingFace on demand.

AIME 24/25GPQAGSM8KHLE LeetCodeDeepWebProgramBench
Quick Start

Up and running in minutes

Set up the environment, then let the MetaAgent run a task to completion.

setup.sh
# 1 · create the Python environment
conda create -n agent python=3.12
conda activate agent
pip install -r scripts/requirements.txt

# 2 · browser automation
pip install playwright && playwright install
pip install browser-use && browser-use install
run.sh
# run the default task
python examples/run_meta_agent.py

# run an inline task
python examples/run_meta_agent.py \
  --task "Reverse a string + add unit tests"

# run from a task document
python examples/run_meta_agent.py \
  --task-file examples/tasks/qsar_egfr_experiment.html

Full setup (Vault secret manager, opencode, Python env) is documented in scripts/INSTALL.md →

Ready to build self-evolving agents?

Clone the repo, run the MetaAgent, and watch the ecosystem grow.