项目 × 研究 × 博客
一张卡片代表一个项目、一篇研究笔记或一篇博客文章。点击浏览详情。
openevolve
开源 AlphaEvolve 风格的进化式编码智能体,使用 LLM 生成程序变体、评估打分并选择高分候选持续迭代。
agents
数据中心化的自进化语言智能体框架,强调用数据、环境反馈和智能体编排让 Agent 在任务中持续改进。
reflexion
Reflexion 经典实现,使用语言反思作为“ verbal reinforcement ”,把失败经验写入记忆以改进后续行为。
AgentEvolver
面向高效自进化智能体系统的研究实现,核心关注轨迹、经验和评估反馈驱动的 Agent 能力提升。
self-refine
Self-Refine 经典实现,用同一个 LLM 生成初稿、生成反馈、再根据反馈迭代精炼,无需训练或强化学习。
SE-Agent
面向代码智能体的自进化框架,通过 Revision、Recombination、Refinement 在多条推理轨迹间交换信息,扩大搜索空间。
science-codeevolve
面向科学与算法发现的 CodeEvolve 实现,将 LLM 与遗传算法、岛屿模型、交叉/变异等机制结合。
SCOPE
Self-evolving Context Optimization via Prompt Evolution,通过提示词/上下文演化优化 LLM 的任务表现。
LLM-Self-Judge
围绕多模态推理中模型自评判和无监督自进化的官方实现,关注 judge 信号如何驱动模型/数据改进。
DARWIN
自进化 LLM 越狱/安全研究框架,维护策略池并用检索、过滤、马尔可夫选择、反思和遗传演化扩展攻击策略。
OPRO
Google DeepMind 的 OPRO(Optimization by PROmpting),用 LLM 作为优化器,通过进化式 Prompt 迭代搜索更优候选。
OpenELM
CarperAI 的开源进化大模型框架,使用 MAP-Elites 质量多样性算法进化 LLM Prompt 和代码。
ADAS
自动设计智能体系统(ADAS),使用 Meta Agent Search 在代码层面搜索候选 Agent 架构。
FunSearch
Google DeepMind 的 FunSearch,用 LLM + 岛屿模型进化搜索发现数学新知识,首个在数学开放问题上超越人类已知的 LLM 系统。
AutoML-Agent
ICML 2025 论文实现,分层管理式 5 Agent 的 AutoML 框架,含 7 状态 FSM + Evol-Instruct 进化 + 代码自动修复。
CoML
微软研究院的交互式 ML 编程助手,通过知识库驱动的配置推荐和 Jupyter Magic 实现经验积累与迁移。
MetaGPT
高可见度的多智能体协作框架,Code = SOP(Team) 理念,含 SELA(MCTS+LLM)和 AFlow(ICLR 2025 Oral)线索。
AutoGPT
GitHub 历史高星标自主 Agent 项目之一,定义了 TAO 循环(Thought-Action-Observation),并演进出平台化形态。
CrewAI
轻量高性能多 Agent 自动化框架,Crew + Flow 双架构,零依赖独立实现,100k+ 认证开发者。
DSPy
斯坦福 NLP 声明式 LLM 编程框架,将 Prompt 视为可优化参数,含 SIMBA 自我反思优化器和丰富 Teleprompter 家族。
AutoGen
微软多智能体对话框架,基于 Actor 模型的消息传递运行时,支持 Round-Robin/Group Chat/Society of Mind 多种编排。
CAMEL-AI
较早探索多 Agent 角色扮演对话的框架,含 Critic 在环批评机制和 Self-Instruct 数据生成。
LangGraph
LangChain 的状态化多 Agent 工作流框架,将 Agent 编排建模为有向图,支持循环/分支/并行和状态持久化。
SWE-Agent
Princeton NLP 软件工程 Agent,ICLR 2025 Oral,提出 Agent-Computer Interface (ACI);SWE-bench 结果需按论文和 leaderboard 版本复查。
OpenHands
开源 AI 软件开发平台(原 OpenDevin),AI Agent + 完整开发环境(编辑器/终端/浏览器),55k+ Stars。
Devika
开源 AI 软件工程师,多 Agent 流水线(规划→研究→决策→编码→测试→修复),支持 8+ LLM 后端。
AgentVerse
清华 OpenBMB 多智能体仿真平台(ACL 2024),双模式(仿真+任务求解),含 Reflection 记忆操作器。
ReEvo
NeurIPS 2024 反射式进化搜索框架,将 LLM 作为语言超启发式(LHHs),通过短期/长期反射机制结合交叉变异实现组合优化算法的自动发现。
LLaMEA
IEEE TEVC 2025 论文,自称 AlphaEvolve 的完全开源后继者,利用 LLM 自动生成和优化元启发式算法,获 GECCO 2025 银质 Humies 奖。
EvoPrompt
ICLR 2024 论文,将进化算法(GA/DE)与 LLM 结合实现离散提示词优化,在 31 个数据集上显著超越人工设计的提示词。
EvoAgent
NeurIPS 2024 论文,通过进化算法自动将专家 Agent 扩展为多智能体系统,模拟人类社会的繁衍机制实现从单 Agent 到多 Agent 的自动化扩展。
EvoAgentX
面向 LLM Agent 的构建、评估与自动进化开源框架,集成 TextGrad/MIPRO/AFlow/EvoPrompt 等多种进化算法优化 workflow 结构与提示词。
EverOS
面向自进化 Agent 的长期记忆平台,把 EverCore、HyperMem、EverMemBench、EvoAgentBench 和 skills/MCP 集成放进同一条记忆到评测的数据流。
OpenTreeSearch
Genentech 发布的基于 PUCT 树搜索的 LLM 引导代码进化工具,构建在 OpenEvolve 之上,用单超参数替代岛屿模型多超参数。
pyribs
USC ICAROS Lab 的质量多样性(QD)优化库,实现高度模块化的 RIBS 框架,支持 CMA-ME/CMA-MEGA/CMA-MAE 等先进 QD 算法。
LLM4EC
IEEE TEVC 综述论文配套资源库,系统整理 LLM 与 EC 交叉领域研究,划分为 LLM 增强 EA、EA 增强 LLM、LLM 与 EA 协同应用三大方向。
LLM4Opt
ACM Computing Surveys 2025 综述配套资源库,提出 LLM 在算法设计中的四范式分类:LLMaO/LLMaP/LLMaE/LLMaD,收录 300+ 篇论文。
Nevergrad
Meta Research 的 Python 无梯度优化平台,实现 100+ 种优化器的统一接口,含差分进化/CMA-ES/粒子群/贝叶斯优化等,通过 NGOpt 自动选择算法。
DEAP
JMLR 2012 经典进化计算框架,支持 GA/GP/ES/多目标/协同进化等多种范式,以算法显式化、数据结构透明化为设计理念,与 multiprocessing/SCOOP 兼容。
pycma
CMA-ES 官方 Python 参考实现,适用于非凸、病态、多模态和噪声连续优化问题,含边界/约束处理、噪声处理和混合整数优化支持。
auto-sklearn
Freiburg 大学 AutoML 团队开发的自动化机器学习框架(NeurIPS 2015),结合元学习、贝叶斯优化和自动模型集成实现端到端 ML 流程自动化。
Self-Evolving-Agents
普林斯顿/UIUC/MIT 等联合撰写的自进化 Agent 综述,系统回答 What/When/How/Where to Evolve 四个核心问题,覆盖从模型权重到 Agent 架构的多个进化层次。
self-improvement-llm
Stony Brook University 发布的 LLM 自改进系统级综述,提出覆盖数据获取/选择/模型优化/推理优化/自主评估的五组件框架,收录 300+ 篇论文。
LLM-EA-Survey
LLM 与 EA 交叉研究综述论文配套资源库,将论文分为 LLM 用于 EA 和 EA 用于 LLM 两大方向,涵盖 2022-2024 年 100+ 篇重要文献。
Tutorial-GP-LLM
MIT ALFA Group 的教学项目,展示如何将 LLM 集成到 GP 进化算子中,扩展 pony_gp 框架实现 LLM 驱动的个体初始化、交叉和变异操作。
LLM-Agent-Optimization
ACM Computing Surveys 接收论文配套列表,系统探讨 LLM Agent 优化技术,划分为参数驱动/无参数/数据集与基准/应用四大类别,收录 200+ 篇论文。
Awesome-Code-LLM
CodeFuse AI (蚂蚁集团) 维护的代码 LLM 全景资源列表,TMLR 论文配套仓库,涵盖模型架构/预训练/代码推理/评估基准等 8 大类别。
AgentBench
清华 THUDM 开发的首个系统性 LLM-as-Agent 评测基准(ICLR 2024),设计 8 个截然不同的评测环境,全面评估规划/推理/工具使用/长程决策能力。
RL4CO
AI4CO 团队的强化学习组合优化基准框架(KDD 2025),基于 PyTorch 生态,包含 20+ 种前沿模型,支持路由/调度/图问题等多种 CO 任务。
Awesome-FM4CO
AI4CO 团队维护的基础模型驱动的组合优化前沿研究索引,收录 2022-2026 年 200+ 篇论文,从 LLM 解决 CO 和领域 FM 两个维度组织文献。
OpenDevin
开源 AI 软件开发平台(原 OpenDevin,现 OpenHands),在 SWE-bench 上达到 77.6% 解决率,被 TikTok/Amazon/Netflix/NVIDIA/Google 等采用。
GP-LLM-Code-Evolution
GP+LLM 融合代码进化综合研究报告,分析遗传编程与大语言模型融合的技术路线,涵盖 LLM 作为遗传算子/优化器/超启发式/进化编码 Agent 等方向。
future-agi
FutureAGI 自改进 AI Agent 框架,支持 Agent 通过反馈循环持续优化自身行为。
awesome-self-evolving-agents
厦门大学 XMUDeepLIT 维护的自进化 Agent 领域论文精选列表,涵盖 Self-Evolving、Self-Improving、Self-Debugging 等方向。
ag2
AG2(原 AutoGen)微软开源的多 Agent 对话框架,支持可组合的 Agent 编排、群聊、人机协作和工具集成。
chatdev
ChatDev 清华 OpenBMB 多智能体协作平台,2.0 版为零代码编排平台,Puppeteer 用 RL 优化 Agent 编排(NeurIPS 2025)。
openagents
OpenAgents 开源 Agent 平台,专注于 LLM 工具使用(Tool Use)和函数调用能力。
superagi
SuperAGI 开源自主 AI Agent 框架,提供丰富的工具生态和 Web UI 管理。
phidata
PhiData Agent 框架,内置记忆系统、知识库和工具集成,支持构建生产级 AI Agent。
smol-developer
Smol Developer 极简 AI 开发助手,以少量代码展示 AI 辅助开发路径。
dify
Dify 开源 LLM 应用开发平台,提供可视化工作流编排、RAG 引擎和 Agent 构建工具。
agentgpt
AgentGPT 浏览器内自主 AI Agent 平台,通过思考→执行→学习循环自动达成目标。
agenta
Agenta 开源 LLMOps 评测平台,集成 Prompt 管理、系统化评估和可观测性。
e2b
E2B 开源 AI 代码安全执行沙箱,基于 Firecracker 微虚拟机实现毫秒级启动的隔离环境。
open-webui
Open WebUI 124K+ stars 自托管 AI 平台,支持 Ollama/OpenAI API、内置 RAG、Python 工具插件。
Gemini CLI Auto Memory
Gemini CLI 是 Google 开源终端 AI Agent;Auto Memory 路线把历史会话挖掘成可审核的 durable memory 和 Agent Skills 草稿。
n8n
n8n 开源工作流自动化平台,支持 400+ 集成和 AI Agent 节点,可视化编排自动化流程。
langflow
Langflow LangChain 生态的可视化 Agent/工作流构建器,拖拽式设计 LangChain 应用。
awesome-agent-papers
Awesome Agent Papers 是 LLM Agent 研究论文精选列表,用于持续追踪 Agent 领域进展。
swe-bench
SWE-bench 软件工程评测基准,使用真实 GitHub Issue 评估 LLM 生成补丁的能力。
osworld
OSWorld 桌面级 OS Agent 评测基准,在真实操作系统环境中评估 AI Agent 的端到端任务完成能力。
webarena
WebArena Web Agent 评测基准,在真实网站环境中评估 Agent 的网页交互任务完成能力。
litellm
LiteLLM 统一 LLM 网关,提供 100+ LLM 提供商的统一接口,支持 Python SDK 和代理服务器。
ollama
Ollama 140K+ stars 本地 LLM 推理运行时,支持 Llama、Gemma、Mistral 等模型的一键部署。
flowise
Flowise 拖拽式 LLM 可视化构建器,通过 UI 拖拽构建 LLM 应用链。
babyagi
BabyAGI 任务驱动自主 Agent,通过目标分解、优先级排序和自我反馈循环自动完成复杂任务。
cheshire-cat
Cheshire Cat AI 模块化聊天框架,通过插件系统实现可扩展的 AI 对话能力。
smolagents
smolagents HuggingFace 轻量级 Agent 框架,简洁 API 构建 AI Agent。
bisheng
Bisheng 开源 LLM 应用开发平台(毕昇),提供可视化编排、知识库管理和 RAG 引擎。
chainlit
Chainlit LLM 聊天应用框架,快速构建类似 ChatGPT 的 LLM 应用界面。
WildClawBench
WildClawBench 是面向 LLM Agent 的真实世界风格评测基准,强调动态、多轮、长程任务与 anti-overfitting 评估设计。
awesome-ai-agents-2026
Awesome AI Agents 2026 精选列表,追踪 2026 年先进 AI Agent 项目和基准。
Awesome Agent Memory by cxxz
A curated agent-memory resource index that helps separate memory tools, patterns and research signals from runtime claims.
Memoir
Memoir frames agent memory as declarative, hierarchical and Git-like, making memory state inspectable through commits and branches instead of only vector retrieval.
Awesome GraphMemory
Awesome GraphMemory is a paper-facing index for graph-based agent memory, giving the corpus a bridge from memory substrates to graph knowledge structures.
ATANT
ATANT adds an agent-continuity evaluation angle: memory is not only retrieval quality, but whether an agent maintains narrative state and checkpoint continuity.
Gitagent
Gitagent treats an agent as a Git repository: identity, rules, memory, tools, skills, hooks, plugins and audit state all live in version-controlled files.
Skillgrade Agent Skill Evaluation
Skillgrade turns Agent Skills into testable assets with repeatable smoke, reliability and regression trials across Gemini, Claude, Codex, ACP and OpenCode-style agents.
Webmaxru Agent Skills
Webmaxru Agent Skills is a reviewed Web API skills collection with explicit creation, validation, remediation and install-check phases.
Waza
Waza is Microsoft's Go CLI / framework for agent skills: it scaffolds skills and eval suites, runs benchmark tasks, compares models, checks coverage, and turns SKILL.md assets into measurable quality gates.
NEXO Brain
NEXO Brain is a local shared-brain memory runtime for Claude Code, Codex, Claude Desktop and MCP clients, combining persistent memory, RAG, temporal recall, forgetting, trust scoring, benchmark reporting and background self-improvement jobs.
state-trace
state-trace is a graph-native memory engine for AI agents that stores typed memories and causal links, then retrieves state-aware context through graph traversal instead of only vector similarity.
Agent Memory Techniques
Agent Memory Techniques is a runnable notebook cookbook for 30 LLM-agent memory patterns, covering short-term buffers, long-term stores, cognitive architectures, multi-agent shared memory, framework integrations, LoCoMo-style evaluation and production patterns.
kbench
kbench normalizes SWE, Terminal-Bench 2.0, tau-bench and Standardized Agent Exams through one CLI and harness contract, including Codex, Claude Code, Gemini CLI, kode-agent-sdk and custom adapter paths.
evmbench
evmbench is a domain-specific benchmark and harness for LLM agents that find and exploit smart-contract bugs, wrapping Codex detect-mode workers, job queues, secret handling, result validation and a report UI.
Skills Best Practices
Skills Best Practices is a methodology layer for professional-grade agent skills, emphasizing discoverable frontmatter, progressive disclosure, deterministic helper scripts and LLM-based validation loops.
SICA Self-Improving Coding Agent
SICA is a self-improving coding-agent implementation where the agent works on its own codebase, making it a core self-referential improvement artifact.
Agent Zero
Agent Zero is an autonomous-agent runtime with Linux execution, memory, browser automation, skills, plugins, projects and subordinate agents.
elizaOS
elizaOS is a high-adoption autonomous-agent framework whose plugin and lifecycle surface can host memory, skill, workflow and evaluator loops.
Centaur
Centaur is a multiplayer self-hosted secure team-agent runtime with Kubernetes sandboxes, bring-your-own harness support, durable workflows, tools and credential boundaries.
Yunjue Agent
Yunjue Agent is a reproducible zero-start in-situ self-evolving agent system that turns execution feedback into reusable tools and auditable traces.
self-evolving-agent
RangeKing/self-evolving-agent is an OpenClaw skill that upgrades self-improvement logs into a governed evolution workspace with curriculum, evals and promotion.
NexAgent
NexAgent uses Elixir/OTP supervision to combine persistent sessions, memory, dynamic tools, skills, cron, subagents and source-level self-improvement.
hermes2anti
hermes2anti is a compact Hermes-inspired memory and skill loop that persists lessons, creates reusable skills, scans them, and recalls prior sessions.
ADHDev
ADHDev is a self-hosted control plane for long-running coding-agent sessions across IDEs and CLIs, including Codex, Claude Code and Hermes Agent.
AI Research SKILLs
AI Research SKILLs is a large open-source skill library for turning coding agents into AI research agents across ideation, experiments, evaluation and paper writing.
ai-skills
ai-skills is a compact behavioral skill library for coding agents, distributing process discipline across Codex, Claude Code, Cursor, Gemini CLI and related runtimes.
Claude Trading Skills
Claude Trading Skills packages trading, DeFi and quantitative finance procedures as reusable Agent Skills for Claude Code, Codex, Cursor, Gemini CLI and other compatible runtimes.
Spec Kit Agent Skills
speckit-agent-skills maps Spec Kit methodology into shared skills, prompts, commands and agent files for Claude Code, Codex CLI, GitHub Copilot CLI, Gemini CLI and related runtimes.
CUGA Agent
CUGA is an enterprise generalist agent harness that wires APIs, MCP servers, policies, memory, knowledge and agent skills into configurable web/API task execution.
AutoR
AutoR is a terminal-first research harness where AI handles execution, humans retain direction, and every run is stored as an inspectable artifact on disk.
Chorus
Chorus is an AI-human collaboration harness that manages session lifecycle, task state, sub-agent orchestration, permissions, observability and failure recovery around LLM agents.
KWeaver Core
KWeaver Core is a harness-first foundation for enterprise decision agents, combining governed context, knowledge networks, tool policies and TraceAI-style feedback evidence.
ClawProBench
ClawProBench is a live-first benchmark harness for evaluating LLM agents in the OpenClaw runtime with deterministic grading, profile selection and repeated-trial reliability.
sd0x-dev-flow
sd0x-dev-flow is a harness layer for Claude Code with dual-review gates, state-machine controls and fail-closed safety for coding-agent workflows.
Utah
Utah is a universally triggered agent harness built on Inngest events with durable retries and local worker execution for think-act-observe loops.
Meta Harness
Meta Harness is a benchmark-driven outer loop for coding-agent harness optimization, with reproducible candidate ledgers and run evidence.
Supermemory
Supermemory 是开源 AI memory 基础设施,把聊天与浏览上下文沉淀为可检索、可复用的长期记忆层。
FlagoS Skills
FlagoS skills 提供开源 AI agent 技能注册表和包格式,把技能资产变成可发布、可安装、可复用模块。
SkillsBench
SkillsBench evaluates how well AI agents actually use reusable skills across specialized multi-step workflows under deterministic and gym-style benchmark settings.
Meta-Harness (Stanford IRIS)
Meta-Harness provides reference code for automated search over model harnesses, including terminal benchmark scaffold evolution and text-classification memory-system search.
Hermes Benchmark Outer-Loop Harness
Hermes Agent Meta-Harness is a standalone outer-loop evaluator that optimizes Hermes benchmark harness candidates without modifying model weights.
SkillX
SkillX is a paper-backed framework that automatically constructs reusable multi-level skill knowledge bases for agents from execution experience.
mem9
mem9 provides a persistent memory server and plugin integrations so multiple agent runtimes can share durable context instead of isolated session notes.
memory-lancedb-pro
memory-lancedb-pro is a production-grade OpenClaw memory plugin that combines long-term storage, hybrid retrieval, and scoped context injection for agent workflows.
GBrain
GBrain is an opinionated long-term brain layer for OpenClaw and Hermes-style agents, combining structured memory pages, graph linking, and retrieval workflows.
Akephalos
Akephalos is a markdown-first local passport for AI agents that carries durable user and project context across different tools and machines.
InternAgent-1.5
InternAgent-1.5 is a unified multi-agent framework for long-horizon autonomous scientific discovery, spanning discovery runs, QA deep research, and reproducible task execution.
ClawXMemory
ClawXMemory is a multi-level OpenClaw memory plugin that writes durable file-based long-term memory and performs model-guided recall for project/user context.
HexAgent
HexAgent is a harness runtime that gives LLM agents an isolated computer layer and pluggable execution protocol instead of coupling agent runtime and machine state.
Agent Harness (EvalOps)
evalops/agent-harness is a lightweight harness adapter that lets one tool registry and one prompt surface run across OpenAI and Claude agent SDK backends.
Harness Evals
Harness Evals is an open-source evaluation framework for LLM agents, prompts, and structured outputs with normalized score semantics.
Browser Harness
Browser Harness connects LLM agents to a real browser through a thin editable CDP harness and explicitly encourages runtime self-healing by letting the agent write missing helper code during execution.
Awesome Agent Skills
junminhong/awesome-agent-skills curates reusable skill workflows and templates for Codex, Claude Code, and other agent platforms.
Trellis
Trellis is an open-source cognitive workspace for autonomous agents with browser execution, persistent workspace memory, and explicit reasoning workflows.
Awesome Agent Harness (Picrew)
Picrew/awesome-agent-harness is a curated index of AI agent harness frameworks, benchmark suites, and reliability-oriented evaluation resources.
Awesome Agent Harness (AutoJunjie)
AutoJunjie/awesome-agent-harness curates the emerging agent harness ecosystem, combining benchmark references and implementation pointers.
Learn Claude Code
learn-claude-code provides a staged curriculum and practical exercises for Claude Code users to build reliable agent coding workflows.
AI Agent Benchmark
ai-agent-benchmark is a compact benchmark repository that evaluates AI agents across coding, math, memory, translation, and safety-oriented task slices.
holaOS
holaOS is an agent environment focused on long-horizon continuity, context retention, and self-evolving workflows for autonomous execution.
Deep Agents
Deep Agents is a batteries-included agent harness that bundles planning, context management, skills, and persistent memory for long-horizon execution.
Archon
Archon is an open-source harness builder for AI coding that turns development processes into deterministic, reusable workflows.
Best of Agent Harnesses
best-of-Agent-Harnesses curates and ranks 100+ agent harness projects, exposing weekly-updated use-case, eval, memory, MCP, and orchestration comparisons for harness selection.
SkillOpt
SkillOpt is a text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts.
TencentDB Agent Memory
TencentDB Agent Memory provides a local-first long-term memory pipeline for AI agents with plugin integration and benchmarked efficiency gains.
GitTaskBench
GitTaskBench is a repository-level benchmark for real-world coding-agent tasks from repository understanding through implementation and task delivery.
OpenClaw ClawHub
ClawHub is the OpenClaw package catalog for discovering, rating, and sharing open-source AI coding agent packages and workflows.
Awesome OpenClaw Skills
awesome-openclaw-skills is a large curated index of OpenClaw skills, prompts, and reusable agent workflow assets.
Awesome OpenClaw
awesome-openclaw curates OpenClaw ecosystem resources including skills, runtime guidance, and adjacent project links.
QuantClaw
QuantClaw is an autonomous quantitative agent runtime built around OpenClaw-style agent workflows for market analysis and execution tasks.
Awesome Scientific Skills
Awesome-Scientific-Skills is a curated list of scientific-agent skills, tools, and workflows intended for reusable research automation.
ClawBench
ClawBench is an open-ended agent benchmark built for evaluating long-horizon OpenClaw-style agent performance and generalization.
Redis Agent Memory Server
agent-memory-server is Redis-maintained memory middleware that gives AI agents fast, persistent, and flexible context memory through an MCP-compatible service and client stack.
LangChain Memory Agent
memory-agent is LangChain’s reference memory-aware agent application that demonstrates persistent user context and long-term adaptive behavior.
iFLYTEK SkillHub
SkillHub is an open-source AI agent platform focused on skill packaging, orchestration, and reusable execution components.
Acontext
Acontext provides an agent memory layer designed to persist and retrieve skill-aware context for autonomous AI workflows.
Nemori
Nemori is an episodic memory and persistent context engine for autonomous agents that need durable recall across sessions.
Continuity Benchmarks
continuity-benchmarks provides reproducible execution-intent memory benchmarks for long-horizon AI coding agents with LongMemEval-S and ID-RAG matrix evaluation.
MemWeave
memweave is a zero-infrastructure async Python memory library for AI agents using markdown files and SQLite-backed searchable persistence.
Procedural Memory Benchmark
Proced_mem_bench benchmarks procedural memory retrieval for language agents across ALFWorld trajectories and upcoming OSWorld scenarios.
RepoMod Bench
mcode-benchmark (RepoMod-Bench) evaluates repository-scale AI agent translation across languages and frameworks with hidden test validation.
CodeScaleBench
CodeScaleBench is a benchmark suite for measuring coding agents with external retrieval tools on large enterprise-scale software tasks.
MLCommons ModelBench
MLCommons modelbench runs safety benchmarks against AI models and publishes detailed hazard-oriented benchmark reports.
Self-Improving Agent
self-improving-agent studies whether coding agents can improve themselves through a repeated eval-and-update loop over real tasks.
Awesome Agent Skills Security
awesome-agent-skills-security curates security and benchmark resources specifically for agent skill systems and tool-use workflows.
Awesome Agent CLIs
awesome-agent-clis tracks command-line agent tools and workflow utilities that can be reused as skill and harness building blocks.
Agent Skills
agent-skills provides reusable operational skill bundles designed to be installed and maintained across agent runtimes.
Agent Memory
agent-memory is an opinionated TypeScript memory runtime focused on extracting, storing, and retrieving long-lived context for AI agents.
Tiermem
Tiermem introduces benchmark tasks that jointly evaluate memory answer quality and provenance consistency for long-term AI assistants.
Awesome Agent Skills (itgoyo)
itgoyo/awesome-agent-skills is a cross-runtime skills index covering Claude Code, Codex, Gemini CLI, Cline, Cursor and MCP-oriented skill resources.
Agent Skill Loader
agent-skill-loader provides a TypeScript loader pipeline for wiring reusable skill modules into AI agent runtimes.
SEAD
SEAD is a benchmark for evaluating LLM-based agents on software system design tasks with an explicit self-evolving-agent framing.
Darwin Mobile Agent
darwin-mobile-agent targets self-evolving mobile automation agents with an explicit iterative improvement workflow.
Mem-Gallery
Mem-Gallery provides a benchmark suite for long-term memory capability in LLM-based agents and assistants.
InfiAgent
InfiAgent presents an open-source framework aimed at self-improving AI agent loops with planner/executor/reflection style workflows.
Clownfish
clownfish is OpenClaw's maintainer codex harness focused on resolving issue clusters at scale with auditable workflows.
Crabpot
crabpot is a compatibility testbed repository for OpenClaw community plugins and plugin seam validation.
Gitcrawl
gitcrawl is a local-first issue and pull-request crawler for creating reusable repository archives used by maintainer and agent workflows.
Crawlkit
crawlkit is OpenClaw's shared Go infrastructure for local-first crawler and archive processing workflows.
Discrawl
discrawl is a Discord CLI crawler with SQLite-backed archival pipelines for building reusable organizational memory datasets.
MemOS Cloud OpenClaw Plugin
MemOS-Cloud-OpenClaw-Plugin is an official OpenClaw plugin that adds hosted long-term memory retrieval and persistence through MemOS Cloud.
OpenAI Swarm
Swarm is OpenAI’s educational multi-agent orchestration framework for lightweight handoffs, routines, and inspectable agent workflow patterns, now explicitly positioned as a predecessor to the OpenAI Agents SDK.
Agent Lightning
Agent Lightning is Microsoft's framework for turning arbitrary agent execution traces into RL-friendly transitions and optimizing agent behavior with LightningRL.
Awesome AI Agent Skills
awesome-ai-agent-skills curates reusable skill packs and installation paths across Codex, Claude, Cursor, Gemini CLI, OpenCode, and related agent runtimes.
ACPX
acpx is OpenClaw's state-preserving runtime for switching across coding agents while retaining full session context and ACP compatibility.
Crabbox
crabbox is OpenClaw's browser-agent benchmark framework for reproducible task execution and evaluator-driven comparison.
OpenClaw Windows Node
openclaw-windows-node provides a Windows companion runtime layer to execute OpenClaw agent workflows in native Windows contexts.
Oh My Agent
oh-my-agent is an open source multi-agent AI framework with workflow orchestration, memory modules, tooling integration, and built-in benchmark surfaces.
ReMe
ReMe is a memory management toolkit for AI agents that provides long-term memory retention, context compression, and benchmark-backed retrieval quality signals.
Open Memory Specification (OMS)
OMS defines a portable and auditable memory standard for autonomous agents, including a binary memory grain format and context assembly language.
MemToMem
memtomem provides a hierarchical long-term memory framework to improve context retention, retrieval quality, and continuity in autonomous agent workflows.
Awesome Agent Skills (Skillmatic)
skillmatic-awesome-agent-skills is a large cross-framework directory of reusable agent skills spanning Codex, Claude Code, Cursor, Gemini, and other runtimes.
Agentic Harness
agentic-harness is a Python package for building and testing AI workflows with composable harness primitives and lightweight runtime checks.
.NET Skills
dotnet/skills curates production-grade coding-agent skill packs with .NET-focused workflows, scoring guidance, and multi-runtime installation support.
Datalayer Agent Skills
datalayer/agent-skills provides composable skill packs and runtime recipes for coding agents with an emphasis on reproducible installation and workflow reuse.
Agent Skills (Block)
block/agent-skills is a public enterprise-oriented skill library that packages repeatable coding-agent playbooks for safer and faster software delivery.
EvalMonkey
EvalMonkey provides a lightweight evaluation harness for LLM agents with regression workflows and reusable benchmark checks.
AEC Bench
aec-bench is an agentic context engineering benchmark suite for measuring long-context agent behavior and retrieval-aware execution quality.
Google AX
google/ax is a production-focused agent runtime framework that packages context engineering, orchestration, and evaluation-friendly primitives for real systems.
AutoHarness
AutoHarness provides a governance-oriented runtime that wraps LLM clients and upgrades demo agents into auditable, policy-controlled production loops.
OWASP Agent Memory Guard
OWASP Agent Memory Guard is a runtime protection layer for agent memory operations, targeting memory poisoning, leakage, and integrity tampering.
Addy Agent Skills
addyosmani/agent-skills is a large-scale production skill pack that standardizes coding-agent engineering workflows and quality gates.
Zo Skills Registry
zocomputer/skills is an agent skills registry that unifies official, external, and community skill packages with validation and manifest generation.
QF-Bench
QF-Bench is a state-aware benchmark dataset for evaluating AI agents on realistic quantitative finance workflows under sandboxed verification.
AegisLLM
AegisLLM is a cooperative multi-agent security framework that claims self-improving defensive behavior through test-time prompt optimization.
EvoMap Evolver
EvoMap/evolver presents a memory-map framework that keeps updating graph structures from user and task feedback to improve retrieval and reasoning quality over time.
Hermes Agent Self-Evolution
NousResearch/hermes-agent-self-evolution packages a self-evolving RL training workflow that distills improved interaction behavior into Hermes model variants.
MCP Neo4j Agent Memory
knowall-ai/mcp-neo4j-agent-memory provides an MCP server that gives agents persistent graph memory backed by Neo4j plus vector retrieval.
CC Harness Skills
LearnPrompt/cc-harness-skills curates reusable harness playbooks for Codex and Claude Code workflows, emphasizing skill reuse over one-off prompts.
ClawArena
aiming-lab/ClawArena focuses on benchmark-centric evaluation for computer-use agents under controlled interactive environments.
Kodu Awesome Agent Skills
kodustech/awesome-agent-skills is a curated index of reusable AI coding-agent skills and prompt workflow patterns.
Lossless Claw Context Management
Martian-Engineering/lossless-claw provides persistent context-management primitives for OpenClaw-style agent workflows with explicit long-horizon memory handling.
AgentsMeetRL Benchmark Index
thinkwee/AgentsMeetRL is an awesome-style collection focused on agentic RL papers, methods, and benchmark references for agent training/evaluation loops.
OpenHarness Agent Harness Fork
syntax-syndicate/OpenHarness-agent-harness tracks an OpenHarness runtime fork for controlled agent execution and evaluation workflow standardization.
OpenClaw Arcade Plugin
ArcadeAI/openclaw-arcade-plugin connects OpenClaw agents to Arcade.dev tool APIs as reusable plugin skills.
ObsidianClaw Knowledge Plugin
humanitylabs-org/obsidianclaw is an archived OpenClaw plugin that links agent workflows with Obsidian-style knowledge memory.
Momo OpenClaw Plugin
momo-personal-assistant/openclaw-plugin provides an OpenClaw plugin layer for personal-assistant style workflows with persistent task context.
NVIDIA Agent Skills Registry
NVIDIA/skills curates reusable agent skill modules and templates for enterprise-grade assistant and coding workflows.
Supabase Agent Skills
supabase/agent-skills publishes composable skill packs for coding and product-support agent workflows with reusable prompt structures.
Skills Vote Evolution Benchmark
MemTensor/skills-vote focuses on self-evolving skill selection and evaluation loops for agent improvement experiments.
Ori Mnemos Memory Harness
aayoawoyemi/ori-mnemos provides a memory substrate that records agent traces and feeds retrieval into future steps.
OpenClaw DingTalk Channel Plugin
soimy/openclaw-channel-dingtalk integrates OpenClaw agents with DingTalk channels for enterprise communication workflows.
OpenClaw Codex App Server
pwrdrvr/openclaw-codex-app-server exposes an app-server runtime linking OpenClaw workflows with Codex and provider integrations.
OpenClaw Observability Plugin
henrikrexed/openclaw-observability-plugin adds observability and trace-monitoring capabilities to OpenClaw agent runtimes.
Self Evolve OpenClaw Playground
longmans/self-evolve is a self-learning OpenClaw plugin that learns from feedback and turns runtime experience into reusable memory.
OpenClaw Langfuse Observability Plugin
MCKRUZ/openclaw-langfuse adds Langfuse observability to OpenClaw through a lightweight plugin bridge.
HyperSpell OpenClaw Memory Engine
hyperspell/hyperspell-openclaw extends OpenClaw with memory/context synchronization and retrieval enhancements.
Awesome OpenClaw Plugins (Composio Community)
composio-community/awesome-openclaw-plugins curates OpenClaw plugin assets and operational playbooks into a structured ecosystem index.
Last30Days Skill Benchmark Harness
mvanhorn/last30days-skill provides a mature Python benchmark harness for evaluating agent skill performance over time.
ClawDreamer Automation Plugin
EESIZ/clawdreamer offers an OpenClaw-oriented automation app/plugin for productivity and workflow execution.
Awesome OpenClaw Plugins (ThisIsJeron)
ThisIsJeron/awesome-openclaw-plugins is a curated OpenClaw plugin knowledge index with operational categories and install references.
Awesome OpenClaw Money Maker Index
BlockRunAI/awesome-OpenClaw-Money-Maker curates OpenClaw-centric automation and skill pathways oriented toward real-world monetization scenarios.
Awesome OpenClaw Ecosystem Index (SamurAIGPT)
SamurAIGPT/awesome-openclaw is a curated index of OpenClaw resources, tools, skills, tutorials, and community projects.
Awesome Claws Collection (LHL3341)
LHL3341/awesome-claws curates OpenClaw products, skills, communities, and ecosystem resources in bilingual scenario-focused structure.
REINS Self-Improving Model Framework
pegasi-ai/reins provides a self-improving control framework aimed at reducing undesired agent behavior during iterative execution.
MentisDB Agent Memory Graph Database
cloudllm-ai/mentisdb is a durable semantic memory engine and versioned skill registry for long-horizon AI agent operation.
Awesome OpenClaw Plugin Index (rohitg00)
rohitg00/awesome-openclaw is a high-activity curated index for OpenClaw plugins, skills, deployment patterns, and ecosystem resources.
ClawMem On-Device Memory Layer
yoloshii/ClawMem provides an on-device memory layer for Claude Code, Hermes, and OpenClaw agents with hybrid retrieval and hooks.
SoulClaw OpenClaw Fork with Soul Memory
clawsouls/soulclaw is an OpenClaw-derived framework emphasizing persistent identity, multi-tier memory, and channel-ready agent runtime.
Commonly Agent Swarm Operating System
Team-Commonly/commonly positions itself as an AI swarm workflow operating system with built-in orchestration and repository-first execution patterns.
MateClaw OpenClaw Memory and Rule Engine
matevip/mateclaw extends the OpenClaw ecosystem with memory-aware runtime controls, rule-engine automation, and agent execution governance.
OnionClaw OPSEC Agent Plugin
christinminor459/OnionClaw is an OPSEC-focused OpenClaw plugin layer for secure agent execution and operational hardening.
Nowledge Community OpenClaw Skills Hub
nowledge-co/community acts as a community hub for OpenClaw skills, runtime integrations, and ecosystem collaboration resources.
LACP Lightweight Agent Context Protocol
0xNyk/lacp provides a lightweight agent context protocol aimed at reusable context passing and tool interoperability.
WATER Self-Improving Coding Agent
manthanguptaa/water targets self-improving coding agents with benchmark-linked execution loops and iterative optimization behavior.
Sponsio Workflow and Multi-Agent Control Plane
SponsioLabs/Sponsio provides workflow-centric multi-agent control infrastructure for orchestrated autonomous task execution.
Agent Workflow Memory Knowledge Graph Runtime
zorazrw/agent-workflow-memory builds an agent workflow memory system with a graph-backed memory manager and FastAPI execution surface.
AgentMemory Benchmark Framework
webzler/agentMemory provides a benchmark framework focused on agent memory capability and hallucination-aware evaluation workflows.
A-MEM Agentic Memory for LLM Agents
agiresearch/A-mem presents agentic memory infrastructure for LLM agents and targets long-term context quality in autonomous task loops.
ElizaOS AgentMemory Plugin
elizaOS/agentmemory is a plugin module that enables memory management capabilities inside ElizaOS-based agent runtimes.
Apify Agent Skills Library
apify/agent-skills curates reusable skills for coding agents to improve execution quality across software and automation tasks.
OpenSwarm Multi-Agent Orchestration Framework
openswarm-ai/openswarm provides lightweight multi-agent orchestration primitives for building swarm-style autonomous workflows.
AgentraLabs Agentic Memory Runtime
agentralabs/agentic-memory provides an agent memory runtime package focused on persistent context and reusable memory APIs.
LHL Agentic Memory Module
lhl/agentic-memory is a simple memory implementation designed to be slotted into any agentic system.
LangChain AgentEvals Harness
langchain-ai/agentevals is a LangChain ecosystem harness for running repeatable agent evaluation suites.
AgentEvals.dev Benchmark Framework
agentevals-dev/agentevals provides an Apache-licensed benchmark framework for evaluating agent behavior and quality changes.
ALucek Agentic Memory Methods Library
ALucek/agentic-memory curates practical memory implementation methods for agentic LLM systems inspired by cognitive architecture concepts.
VectorSpaceLab General Agentic Memory
VectorSpaceLab/general-agentic-memory focuses on generalized memory abstractions that can be reused across varied agent workloads.
Nocturne Memory Context Engine
Dataojitori/nocturne_memory is a context-aware long-term memory engine for AI agents with sustained repository activity.
Plaited Agent Eval Harness
plaited/agent-eval-harness provides a Unix-style, schema-driven evaluation harness for CLI agents with run, grade, compare, and calibrate modes.
MultiAgentEval Enterprise Harness
najeed/ai-agent-eval-harness is an enterprise-grade reliability framework for AI agents with benchmark, replay, and verification surfaces.
Arc Continual Learning Benchmark
Arc-Computer/CL-Bench is a benchmark framework for evaluating LLM agent continual learning in stateful environments with CRM-style workflows.
ClawVault Memory Runtime
Versatly/clawvault provides structured persistent memory for AI agents and exposes benchmarks, docs, and eval surfaces around that memory stack.
OpenSquilla Token-Efficient Agent Runtime
opensquilla/opensquilla is a token-efficient AI agent runtime focused on higher intelligence density with OpenClaw, memory, and MCP integration signals.
BitFun Desktop Agent Runtime Suite
GCWing/BitFun provides a desktop-grade agent runtime suite with built-in code/cowork/computer-use flows plus persistent memory and evolving behavior claims.
PowerMem Agent Memory Plugin
oceanbase/powermem is an agent memory plugin focused on improving accuracy, agility, and affordability for AI agent memory retrieval.
Mnemon Persistent Memory Substrate
mnemon-dev/mnemon provides LLM-supervised persistent graph memory for agents across Claude Code, OpenClaw, and other CLI runtimes.
Mengram Human-Like Agent Memory
alibaizhanov/mengram offers human-like semantic/episodic/procedural memory infrastructure for AI agents with multi-framework integrations.
IMCodes Shared Agent Context Layer
im4codes/imcodes focuses on shared agent context and memory with supervised execution and cross-agent audit capabilities.
Weaviate Query Agent Benchmarking Toolkit
weaviate/query-agent-benchmarking provides benchmarking scenarios and tooling for evaluating Weaviate query-agent behavior.
PinchBench Skill
PinchBench is an OpenClaw-oriented real task benchmark that measures coding agents across productivity, research, writing, coding, analysis, email, memory, and skill-discovery tasks.
Agent Memory Benchmark
vectorize-io/agent-memory-benchmark is a benchmark harness for comparing agent memory strategies on multi-step tasks where retrieval quality, latency, and token cost all matter.
EvoMap Awesome Agent Evolution
EvoMap/awesome-agent-evolution is a curated awesome list covering self-evolution, memory, protocols, benchmarks, coding agents, and adjacent infrastructure.
RepairAgent
RepairAgent is an autonomous agent that fixes bugs in Java projects using LLMs through a localize-analyze-fix-test iteration loop.
AccelOpt
AccelOpt is a self-improving LLM agentic system that iteratively optimizes AI accelerator kernels using optimization memory and benchmarked kernel profiling.
DM-Code-Agent
DM-Code-Agent is a local-first and auditable Python code agent baseline with explicit planning, trace replay, optional reflexion modules, and benchmark-facing evaluation paths.
LaSeR
LaSeR is a reinforcement-learning recipe that jointly improves reasoning and self-rewarding behavior by adding an MSE self-reward term to the RLVR objective.
GraphMind
GraphMind is an agentic RAG runtime that combines knowledge graphs, dual orchestration engines, and self-evaluating retrieval pipelines for autonomous knowledge work.
Awesome Agent Memory
Awesome Agent Memory is an up-to-date survey index for agent-memory papers, systems, and benchmarks rather than a runnable memory runtime.
Learning SDK
Learning SDK is a drop-in memory and continual-learning substrate for existing LLM agents, exposing a thin wrapper instead of a full custom runtime.
Awesome Code as Agent Harness Papers
Awesome Code as Agent Harness Papers is a harness-first survey index that turns code-centric agent systems into an organized reading and comparison map.
Geneclaw
Geneclaw is a safe-by-default self-evolving agent framework that observes failures, proposes fixes, and only applies changes behind an explicit safety gate.
ClawBench
ClawBench is a benchmark for agent systems that scores the full stack through execution traces, reliability metrics, and diagnostics rather than only final-task success.
Awesome Agent Skills
voltagent/awesome-agent-skills is a high-signal resource index for reusable agent skills across Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, and adjacent coding-agent surfaces.
swarmclaw
swarmclawai/swarmclaw is a self-hosted multi-agent runtime that combines memory, MCP tooling, delegation, schedules, skills, and swarm-oriented execution surfaces.
Swarms
Swarms is a production-oriented multi-agent orchestration framework that packages sequential, concurrent, and hierarchical agent topologies with tools, memory, protocol adapters, and marketplace-style extensions.
Agent Orchestrator
Agent Orchestrator is a production coding-agent swarm that runs spec, build, test, review, and merge flows through Git worktrees, skills, memory, and MCP-connected execution.
MedMemoryBench
MedMemoryBench is a benchmark for agent memory in personalized healthcare, testing whether clinical agents retain, retrieve, and use longitudinal patient context safely and accurately.
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.
Agency Swarm
Agency Swarm is a mature Python orchestration framework that layers role-aware communication flows, type-safe tools, thread persistence, and terminal/web demos on top of the OpenAI Agents SDK.
XSkill
XSkill is a multimodal continual-learning paper implementation that extracts task-level skills and action-level experiences from agent trajectories, stores them in a memory bank, and re-injects them during inference across benchmark suites.
LightAgent
LightAgent is a lightweight Python agent framework that combines persistent memory, MCP integration, native skills, LightSwarm collaboration, and newly surfaced LightFlow workflow orchestration into a small-footprint self-learning runtime.
OpenAI Agents SDK (Python)
OpenAI Agents SDK (Python) is the official Python baseline for multi-agent orchestration with handoffs, tool calls, guardrails, tracing, and session state.
OpenAI Agents SDK (JS)
OpenAI Agents SDK (JS) is the official TypeScript baseline for building agent orchestration loops with tools, handoffs, guardrails, realtime hooks, and tracing.
Agentic Harness Engineering
Harness engineering is the layer where prompts, tools, middleware, memory, subagents, and evaluators become editable agent infrastructure rather than fixed wrappers around a base model.
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.
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.
OpenClaw
OpenClaw is a local-first personal AI agent runtime with multi-channel interaction, durable sessions, skills, companion apps, and product-facing operating workflows.
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.
agentmemory
agentmemory is a cross-harness persistent memory layer for coding agents and MCP clients, with wiki-style knowledge capture, retrieval, and continuity signals.
Self-Refine:基于自反馈的迭代式优化
同一 LLM 生成答案、批评答案并在推理期迭代改写,是轻量 Self Evolve 循环的基线模式。
Reflexion:带语言强化学习的智能体
把失败反馈转化为自然语言反思并写入记忆,让后续任务在不更新权重的情况下改进。
Teaching LLMs to Self-Debug
把执行结果、测试和调试解释作为程序修复信号,直接支撑代码智能体的回归门禁。
SPIN:Self-Play Fine-Tuning
通过自博弈微调把弱模型推向更强模型,把自我改进从提示期推进到训练期。
OPRO:Large Language Models as Optimizers
用自然语言描述目标和历史分数,让 LLM 基于反馈提出新候选,是通用“LLM 即优化器”模板。
FunSearch:LLM + 评估器的程序进化
把 LLM 与自动评估器组合,进化数学和科学发现程序,是“生成器 + 验证器 + 谱系”的重要案例。
AutoML-Agent:多 Agent 全流程 AutoML
将数据检索、预处理、建模、评估和部署拆成专用 Agent,最接近产品化 Self Evolve 工作流。
EvoPrompting:代码级神经架构搜索
把语言模型作为架构代码的变异算子,用 Benchmark 选择候选,体现“架构即基因组”。
AutoGen / MetaGPT / CrewAI:Agent 编排底座
Agent 框架提供角色、工具和流程编排;Self Evolve 可在其上叠加评估器、记忆、谱系和回归防护。
Agent Symbolic Learning:符号反向传播驱动的自进化智能体
将智能体视为符号网络,用"语言梯度"在自然语言空间做反向传播,让提示词、工具和管线自动进化。NeurIPS 2024。
Darwin Gödel Machine:开放式自进化 Agent 归档
结合达尔文进化与哥德尔机自引用,维护开放式 Agent 归档库。SWE-bench 20%→50%,Polyglot 14.2%→30.7%。
Gödel Agent:运行时 monkey patching 自修改
通过 monkey patching 实现 Agent 在运行时的自修改,结合自我评估与进化策略动态调整自身代码。
ADAS:智能体系统自动设计
在图灵完备的 Python 代码空间中搜索最优 Agent 架构,发现的 Agent 跨领域、跨模型迁移且超越人工设计 SOTA。ICLR 2025。
AlphaEvolve:Gemini 驱动的进化式算法发现
结合 Gemini Flash(广度)+ Pro(深度)做 MAP-Elites 质量多样性搜索;报告了 4×4 complex-valued matrix multiplication 的 48-scalar result。Google DeepMind。
Absolute Zero:零数据自博弈推理强化学习
单一模型同时提出任务并求解,仅用自生成数据进行自博弈 RL 训练,彻底消除对外部标注数据的依赖。NeurIPS 2025。
SelfEvolve:自生成知识 + 迭代自调试
LLM 自生成 API 文档作为知识来源,然后在沙盒中执行代码并将错误反馈回 LLM 进行迭代修复。
RISE:推理期自编辑强化学习
通过迭代式自编辑(self-edit)让 LLM 在推理时修改自身输出,结合 RL 训练让模型学会何时及如何修改。
RAGEN:轨迹级 Agent 强化学习
在轨迹级别对 Agent 进行强化学习训练,让 Agent 从完整交互历史中学习策略改进。
ReVeal:多轮代码 RL 持续改进
多轮代码生成与验证的 RL 框架,让模型在迭代修复循环中持续改进代码质量。Google DeepMind。
Anthropic Dynamic Workflows:为什么它是 Agent-Swarm Evolve 的热点证据
把 Claude Code dynamic workflows 放回 Self Evolve 主题:它不是普通多 Agent,而是让任务组织、验证队列、并行子代理和安全边界一起变成可审计的 Agent-Swarm Evolve。
Anthropic 五月热点:从 Opus 4.8、Stainless 到 965B 估值,AgentOps 栈正在成型
把 Anthropic 2026 年 5 月的 Opus 4.8、Dynamic Workflows、Stainless 收购、Claude containment 和 Series H 融资放回 AI Agent 自进化主题:模型竞争正在变成 AgentOps 基础设施竞争。
Agent 框架不是自进化:AutoGPT、MetaGPT、AutoGen、CrewAI、DSPy、LangGraph 差在哪
从英文论文第六章拆解主流 Agent 框架:它们提供运行时、角色、对话、流程、prompt 编译或状态图,但自进化需要额外的评估、记忆、更新和治理层。
自进化 Agent 怎么评估:别只看分数,要看改进是否可复现
从英文论文第五章拆解自进化 AI 的评估问题:代码、数学、Agent、开放式 benchmark、Star 传播信号、过程指标和推荐评估协议。
进化式代码与算法发现:为什么 AlphaEvolve、DGM 和 OpenEvolve 重要
从英文论文第四章拆解进化式代码和算法发现:LLM 作为优化器、语义变异器、程序搜索器,以及 AlphaEvolve、Darwin Gödel Machine、OpenEvolve 的工程意义。
自进化 Agent 的五个循环:反馈、搜索、评估、反思与种群
从英文论文第二章拆解 Self Evolve 的 Five Evolution Loops:Specification-to-Execution、Search、Evaluator、Reflection、Population,以及如何组合成真实 Agent 系统。
自进化 AI 的未来路线图:评估器、记忆、安全、组合性和生产治理
从英文论文第八章拆解未来方向:评估瓶颈、长期记忆漂移、安全与对齐、五大循环组合、生产挑战和 Self Evolve 两到三年路线图。
AI Agent 自进化入门:从静态模型到会改进自己的系统
把 Self-Evolving AI Agents 英文论文第一章拆成一篇可读博客:解释什么是 AI 自进化、它和在线学习/AutoML/普通 Agent 的边界,以及为什么真正的问题是可验证的自我修改。
LLM 自我改进方法全图:Self-Refine、Reflexion、RISE 到 Absolute Zero
从英文论文第三章拆解 LLM/Agent 自我改进方法:推理时修正、反思记忆、训练期自博弈、RL、语言梯度,以及什么时候该用 prompt、fine-tuning 或 reinforcement learning。
用户真正痛的不是 Agent 不够聪明,而是不可靠、不可观测、不可控
从英文论文第七章拆解自进化 Agent 的用户痛点:幻觉、循环、工具误用、上下文溢出、调试困难、成本、部署、监控与治理。
ADAS 到 DGM:智能体架构自动搜索的进化之路
追踪从 ADAS(ICLR 2025)到 Darwin Gödel Machine 的技术演进:从图灵完备搜索到开放式进化归档。UBC Jeff Clune 团队如何让 Agent 自己设计自己。
Reflexion 深度解读:用语言反思替代参数更新
Reflexion 如何把任务失败转化为自然语言记忆,让 Agent 在不更新权重的情况下持续改进。
AI 自我进化:从概念到实现的证据导读
什么是 AI Self Evolution?本文从 Self Evolve 视角梳理自我进化智能体的核心概念、技术路线与实现路径,帮助读者先建立可复核判断框架。
什么是 AI 自我进化?一张面向开发者的地图
用工程视角梳理自进化智能体如何改进提示词、工具、记忆、代码、工作流与策略,并给出可变对象、反馈信号、验证门和回滚边界。
AlphaEvolve 深度解析:从 MAP-Elites 到算法发现证据
分析 Google DeepMind AlphaEvolve 的双模型架构、MAP-Elites 质量多样性搜索和 4×4 complex-valued matrix multiplication 的 48-scalar result,并标出复核边界。
代码自我改进:从智能体补丁到回归门禁
一套轻量实践流程:让编码智能体自调试、写测试、修复失败并重新验证,同时保留基线、回归切片、失败记录和可审计改进证据。
AI 自我进化的跨领域图谱:从 AutoML 到 Agent 自修改
梳理 AutoML/NAS、进化计算、LLM 自我改进与 Agent 框架之间的深层连接,揭示 Self Evolve 的技术全景。
进化计算 × 大语言模型:2024-2025 前沿导读
导读进化计算与 LLM 融合的代表性进展:从 OPRO、FunSearch 到 AlphaEvolve,从 LLaMEA 到 OpenEvolve。本文是阅读入口,不是系统综述协议。
如何构建自我进化的 AI Agent:最小闭环设计笔记
面向开发者的 Self-Evolving Agent 构建笔记。从最小闭环开始,逐步添加评估器、记忆、进化策略和回归防护,并标出哪些环节仍需复核。
LLM 即优化器:从 OPRO 到 FunSearch 的进化式编码
解读 OPRO、FunSearch、ReEvo 和 LLaMEA 四篇核心论文,理解 LLM 如何充当变异器、重组器和候选生成器。
LLM 自我改进的五种范式:从 Self-Refine 到 Absolute Zero
对比分析 Self-Refine、Reflexion、Agent Symbolic Learning、RISE 和 Absolute Zero 五种 LLM 自我改进范式的方法、优劣与适用场景。Self Evolve 核心技术解读。
开源自进化 AI 项目巡礼:10 个可复核的仓库
从 Self Evolve 项目索引中选取 10 个覆盖不同机制的开源自进化 AI 项目,涵盖进化式代码优化、Agent 进化框架、反思记忆与自评判训练。本文说明它们适合回答什么问题,也标出证据边界。
Self Evolve 研究者图谱:谁在推动 AI 自我进化
梳理 AI 自我进化领域的核心研究者、实验室与合作网络,从 Jeff Clune 的开放式进化到 Google DeepMind 的 AlphaEvolve。
Self Evolve GitHub 项目索引如何阅读
从 OpenEvolve、AgentEvolver、Reflexion 到 Self-Refine,理解自进化 AI 项目的四类实现路线。
SelfEvolve 论文深度解读:LLM 如何实现代码自我改进
详细解读 SelfEvolve 论文的方法、公式与实验结果。探索 LLM 自生成知识与迭代自调试如何实现代码自我改进,以及与 Self Evolve 生态的关联。