What changes?
Look for the object that can change: prompt, workflow, memory, evaluator, code, architecture, or agent organization.
Projects / English mirror
Use the project index as an evidence map, not a leaderboard.
The repository list answers one practical question: which projects are worth inspecting next, and why? It connects each project to a mechanism label, model-card route, source report, visible GitHub signals, and a non-reproduction boundary.
686classified GitHub repositories
300project cards on this site
2,853,793historical adoption prior
How to read
Four checks before citing a project.
Who verifies it?
A repository is stronger when it has tests, benchmarks, reviewer feedback, or reproducible run logs.
Does the change persist?
Retained changes matter more than one-off demos. Check whether the system keeps a better artifact or policy.
What is still unproven?
Many reports are source-tracing notes. Treat stars and descriptions as visible signals, not proof of quality.
Mechanism groups
The useful unit is a mechanism family, not a single famous repo.
The full project index groups repositories by role: algorithm/code evolution, self-reflection, memory or skill substrate, evaluation harnesses, and multi-agent organization. This keeps the page teachable: compare projects that claim the same kind of improvement before trusting the strongest claim.
4 repos
进化式代码 / AlphaEvolve 类
algorithmicsuperintelligence/openevolve, inter-co/science-codeevolve, JARVIS-Xs/SE-Agent, zhang677/accelopt
3 repos
LLM 作为优化器
google-deepmind/opro, carperai/openelm, google-deepmind/funsearch
1 repos
Agent 架构自动搜索
ShengranHu/ADAS
4 repos
Agent 自进化系统
modelscope/AgentEvolver, aiwaves-cn/agents, JarvisPei/SCOPE, clawland-ai/geneclaw
2 repos
反思 / 精炼经典范式
noahshinn/reflexion, madaan/self-refine
3 repos
安全、评判与数据/模型自进化
ZJU-LLM-Safety/DARWIN, OPPO-Mente-Lab/LLM-Self-Judge, rucbm/laser
Recent signals
Recent activity is a signal; it is not proof.
Personal agent product runtime
OpenClaw
OpenClaw is a local-first personal AI agent runtime with multi-channel interaction, durable sessions, skills, companion apps, and product-facing operating workflows.
Self-improving personal agent runtime
Hermes Agent
Hermes Agent is a self-improving personal-agent runtime that combines memory, skill creation, session search, user modeling, messaging gateways, and scheduled automations in one user-facing system.
Compounding Lead-Worker Agent Runtime
Agent Swarm
Agent Swarm is a compounding lead-worker runtime that packages delegation topology, worker isolation, shared memory, MCP tooling, and recurring workflows into one public swarm system.
Agentic skills methodology
Superpowers
Superpowers is a cross-agent skill and workflow methodology library that packages reusable operating habits for Claude Code, Codex, Cursor, and adjacent coding agents.
Token-efficient self-evolving agent
GenericAgent
GenericAgent is a minimal self-evolving autonomous agent framework that grows a skill tree from a small seed loop instead of preloading a large fixed procedure library.