Awesome Agent Evolution

AI Agent self-evolution, memory systems, autonomous self-improvement, and the infrastructure that powers them.

Contents

• Taxonomy
• Agent Evolution and Self-Improvement
• Memory Systems
• Agent-to-Agent Protocols
• Agent Development Platforms
• Agent Coding and Software Engineering
• Prompt and Behaviour Optimization
• Agent Safety and Guardrails
• Embodied AI
• Key Research Papers
• Benchmarks and Evaluation
• Community and Knowledge

Taxonomy

graph LR
    Root["Agent Evolution"] --> SingleAgent["Single-Agent<br/>Optimization"]
    Root --> Infrastructure["Infrastructure<br/>& Protocols"]

    SingleAgent --> Evolution["Self-Evolution<br/>& Self-Improvement"]
    SingleAgent --> Memory["Memory<br/>Systems"]
    SingleAgent --> PromptOpt["Prompt & Behaviour<br/>Optimization"]

    Infrastructure --> Protocols["A2A & MCP<br/>Protocols"]
    Infrastructure --> Platforms["Agent Development<br/>Platforms"]
    Infrastructure --> Coding["Agent Coding &<br/>Software Engineering"]
    Infrastructure --> Safety["Agent Safety<br/>& Guardrails"]
    Infrastructure --> Embodied["Embodied AI<br/>& Device Control"]

Loading

Agent Evolution and Self-Improvement

Projects focused on enabling AI agents to evolve, learn, and improve autonomously.

• Eliza - Autonomous agents for everyone. A framework for creating and deploying AI agents that evolve over time. by @elizaOS (18,128 stars)
• SuperAGI - A dev-first open source autonomous AI agent framework. Build, manage and run self-improving autonomous agents. by @TransformerOptimus (17,431 stars)
• Agent Zero - General-purpose AI agent framework that learns and evolves through interaction. by @agent0ai (16,839 stars)
• Agents (aiwaves) - An open-source framework for data-centric, self-evolving autonomous language agents. by @aiwaves-cn (5,897 stars)
• OpenEvolve - Open-source evolutionary coding agent inspired by AlphaEvolve. Evolves code solutions through LLM-driven mutation and selection. by @codelion (5,893 stars)
• EvoAgentX - Automated framework for evolving agentic workflows. Optimizes agent prompts, tools, and pipelines via evolutionary algorithms. by @EvoAgentX (2,710 stars)
• HyperAgents - Self-referential self-improving agents by Meta. DGM-Hyperagents add an optimization layer so agents edit their own improvement process. by @facebookresearch (2,176 stars)
• evolver - The GEP-powered self-evolution engine for AI agents. Genome Evolution Protocol enables agents to evolve autonomously via mutation and selection. by @EvoMap (1,835 stars)
• Agent0 - Self-evolving agent framework from UNC/Salesforce/Stanford. Improves without human-curated datasets via curriculum and executor agent competition. by @aiming-lab (1,137 stars)
• Darwin Godel Machine - Open-ended evolution of self-improving agents. Agents that rewrite their own code to improve performance. by @jennyzzt (800 stars)
• Ouroboros - Self-creating AI agent that writes its own code and evolves autonomously. Completed 30+ evolution cycles in first 24 hours with zero human intervention. by @razzant (480 stars)
• A-Evolve - The PyTorch for Agentic AI. Open-source infrastructure that evolves any agent across any domain with zero human intervention. #1 on MCP-Atlas (79.4%). by @A-EVO-Lab (451 stars)
• Alita - Generalist Agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution. by @alita-ai (400 stars)
• SEAgent - Self-Evolving Computer Use Agent with Autonomous Learning from Experience. by @SunzeY (238 stars)

Memory Systems

Vector, graph, episodic, and hybrid memory architectures for persistent agent cognition.

• Mem0 - Production-ready AI agent memory with scalable long-term memory. 26% improvement over baseline on LOCOMO benchmark with 91% latency reduction. by @mem0ai (52,250 stars)
• Letta - Platform for building stateful agents with advanced self-editing memory. Formerly MemGPT. by @letta-ai (21,938 stars)
• Cognee - Knowledge engine for AI agent memory. Build and query knowledge graphs from unstructured data in 6 lines of code. by @topoteretes (15,022 stars)
• Memvid - Single-file memory layer for AI Agents in Rust. +35% SOTA on LoCoMo with ultra-low latency (0.025ms P50). by @memvid (14,693 stars)
• memU - Memory system for 24/7 proactive agents. Persistent memory across sessions and platforms. by @NevaMind-AI (13,331 stars)
• MemMachine - Universal memory layer for AI agents. Episodic (graph-based), profile (SQL), and working memory with scalable storage and retrieval. by @MemMachine (5,361 stars)
• EverMemOS - Long-term memory for 24/7 AI agents across LLMs and platforms. by @EverMind-AI (3,674 stars)
• Acontext - Open-source skill memory layer for AI agents. Automatically captures learnings from agent runs and stores them as reusable skill files. by @memodb-io (3,283 stars)
• ReMe - Memory management kit for agents. File-based and vector-based memory systems. SOTA on LoCoMo and HaluMem benchmarks. by @agentscope-ai (2,639 stars)
• mcp-memory-service - Open-source persistent memory for AI agent pipelines. REST API + knowledge graph + autonomous consolidation. by @doobidoo (1,624 stars)
• A-MEM - Agentic Memory for LLM Agents. Self-organizing memory that autonomously manages what to remember and forget. by @agentic-memory (1,500 stars)
• nocturne_memory - Lightweight, rollbackable Long-Term Memory Server for MCP Agents with graph-like structured memory. by @Dataojitori (922 stars)
• Mem9 - Unlimited persistent memory layer for AI agents. Cloud-synced memory across sessions and tools. by @mem9-ai (912 stars)
• Awesome-AI-Memory - Curated knowledge base on AI memory for LLMs and agents, covering long-term memory, reasoning, retrieval, and system design. by @IAAR-Shanghai (686 stars)
• TeleMem - High-performance drop-in Mem0 replacement. 19% higher accuracy, 43% fewer tokens, and 2.1x speedup via narrative dynamic extraction. by @TeleAI-UAGI (457 stars)
• MemSkill - Learning and evolving memory skills for self-evolving agents. Meta-memory that determines what to extract, remember, and forget. by @ViktorAxelsen (398 stars)
• Awesome-Agent-Memory - Curated systems, benchmarks, and papers on memory for LLMs/MLLMs -- long-term context, retrieval, and reasoning. by @TeleAI-UAGI (330 stars)

Agent-to-Agent Protocols

Standards and protocols for inter-agent communication and interoperability.

• Google A2A - Google's open Agent-to-Agent protocol. Enables agent discovery, secure collaboration, and long-running tasks while preserving agent opacity. by @google (23,069 stars)
• mcp-use - The fullstack MCP framework to develop MCP Apps for ChatGPT/Claude and MCP Servers for AI Agents. by @mcp-use (9,692 stars)
• GEP MCP Server - MCP Server for Genome Evolution Protocol. Exposes evolution tools to Claude Desktop, Cursor, and any MCP client. by @EvoMap (1 stars)

Agent Development Platforms

Platforms and tools for building, deploying, and managing AI agents.

• dify - Production-ready platform for building agentic AI workflows with visual orchestration. by @langgenius (136,670 stars)
• LangChain - Full-stack agent engineering platform with composable chains, tools, and memory integration. by @langchain-ai (132,743 stars)
• OpenHands - Open platform for AI software developers as generalist agents. Autonomous coding, debugging, and deployment. by @All-Hands-AI (70,794 stars)
• langgraph - Build resilient language agents as stateful graphs with persistence and streaming. by @langchain-ai (28,668 stars)
• AgenticSeek - Fully local autonomous agent with browsing, coding, and multi-agent capabilities. No API keys required. by @Fosowl (25,849 stars)
• haystack - Open-source AI orchestration framework for building context-engineered production applications. by @deepset-ai (24,758 stars)
• mastra - TypeScript framework for building AI-powered applications with agent workflows and RAG. by @mastra-ai (22,783 stars)
• Coze Studio - AI agent development platform with visual tools for creating, debugging, and deploying agents. by @coze-dev (20,436 stars)
• Google ADK - Open-source Python toolkit by Google for building, evaluating, and deploying sophisticated AI agents. by @google (18,806 stars)
• Parlant - The conversational control layer for customer-facing AI agents. A context-engineering framework for controlling interactions. by @emcie-co (17,876 stars)
• OpenFang - Open-source Agent Operating System for deploying and managing AI agents. by @RightNow-AI (16,435 stars)
• PydanticAI - Type-safe AI agent framework built on Pydantic with structured outputs and dependency injection. by @pydantic (16,165 stars)
• CoPaw - Co Personal Agent Workstation built on AgentScope. Desktop agent platform with multi-agent collaboration and tool integration. by @agentscope-ai (14,779 stars)
• ten-framework - Open-source framework for building conversational voice AI agents. by @TEN-framework (10,403 stars)
• agents - Framework for building realtime voice AI agents with speech-to-speech pipelines. by @livekit (9,963 stars)
• MS-Agent - Lightweight framework by ModelScope to empower agentic execution of complex tasks with memory and deep research. by @modelscope (4,140 stars)

Agent Coding and Software Engineering

AI agents that write, debug, and maintain code autonomously.

• Claude Code - Terminal-native agentic coding tool from Anthropic. Understands your codebase and executes tasks through natural language. by @anthropics (110,814 stars)
• Codex - Lightweight coding agent from OpenAI written in Rust. Runs locally as CLI, IDE extension, or desktop app. by @OpenAI (73,799 stars)
• Aider - AI pair programming in your terminal. Edit code with LLMs across 100+ languages with deep Git integration. by @Aider-AI (42,981 stars)
• Goose - Open-source extensible AI agent that installs, executes, edits, and debugs code autonomously. by @block (39,209 stars)
• goose - Open-source extensible AI coding agent that goes beyond code suggestions. by @aaif-goose (39,209 stars)
• Roo Code - AI coding agent providing a full dev team of specialized agents inside your code editor. by @RooCodeInc (23,025 stars)
• Devika - The first open-source implementation of an Agentic Software Engineer. An open-source alternative to Devin. by @stitionai (19,498 stars)
• SWE-Agent - Automatically fix GitHub issues and handle cybersecurity challenges. State-of-the-art on SWE-bench. by @SWE-agent (18,947 stars)
• Plandex - Open-source AI coding agent designed for large projects and complex real-world tasks with persistent context. by @plandex-ai (15,216 stars)
• Trae Agent - LLM-based agent by ByteDance for general-purpose software engineering tasks. by @bytedance (11,277 stars)
• Open SWE - Open-source asynchronous coding agent by LangChain for software engineering tasks. by @langchain-ai (9,301 stars)
• agent-skills - Production-grade engineering skills and best practices for AI coding agents. by @addyosmani (8,703 stars)
• Mini-SWE-Agent - The 100-line AI agent that solves GitHub issues. Radically simple but scores >74% on SWE-bench verified. by @SWE-agent (3,710 stars)
• Reflexion - Language agents with verbal reinforcement learning. Agents that learn from mistakes through self-reflection. by @noahshinn (3,115 stars)

Prompt and Behaviour Optimization

Tools and frameworks for automatically optimizing agent prompts, instructions, and behavioral patterns.

• Promptfoo - Open-source LLM evaluation and red-teaming framework. Test prompts, agents, and RAGs with 90+ model providers and 67+ security plugins. by @promptfoo (19,731 stars)
• TextGrad - Automatic differentiation via text. Backpropagation through LLM-provided textual gradients, published in Nature. by @zou-group (3,474 stars)

Agent Safety and Guardrails

Projects focused on controlling agent actions, enforcing policies, and preventing harmful behavior.

• NeMo Guardrails - NVIDIA's toolkit for adding programmable guardrails to LLM conversational systems. Policy-based safety controls. by @NVIDIA (5,937 stars)

Embodied AI

Projects connecting AI agents to physical devices, robotics, and real-world environments.

• Open-AutoGLM - An Open Phone Agent Model and Framework. Unlocking the AI Phone for Everyone. by @zai-org (24,750 stars)
• LeRobot - Open-source robotics framework by Hugging Face. Models, datasets, and tools for real-world robotics in PyTorch. (23,031 stars)
• Nanobrowser - Chrome extension for AI-powered web automation. Run multi-agent workflows using your own AI keys. by @nanobrowser (12,647 stars)
• XcodeBuildMCP - A MCP server and CLI for agent use when working on iOS and macOS projects. by @getsentry (5,074 stars)
• Mobile MCP - Model Context Protocol Server for Mobile Automation and Scraping (iOS, Android, Emulators and Real Devices). by @mobile-next (4,420 stars)
• ROS-LLM - Framework for embodied intelligence in ROS. Natural language interactions with LLMs for robot control. by @Auromix (766 stars)

Key Research Papers

Surveys

• A Comprehensive Survey of Self-Evolving AI Agents (arXiv'25) - Unified framework with four components: System Inputs, Agent System, Environment, and Optimisers. Covers evolution of models, prompts, memory, tools, and workflows.
• A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve (TMLR'26) - Organizes around what to evolve, when to evolve, and how to evolve. Covers intra-test-time and inter-test-time adaptation.
• Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers (arXiv'26) - Formalizes agent memory as write-manage-read loop. Taxonomy spanning temporal scope, representational substrate, and control policy.

Self-Evolution and Lifelong Learning

• Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly? (arXiv'25) - First live agent that autonomously evolves itself during runtime. 77.4% on SWE-bench Verified.
• EvoClaw: Evaluating AI Agents on Continuous Software Evolution (arXiv'26) - Benchmark revealing performance drops from >80% to at most 38% in continuous evolution settings.
• Symbolic Learning Enables Self-Evolving Agents (arXiv'24) - Agents that evolve through symbolic representation learning.
• Building Self-Evolving Agents via Experience-Driven Lifelong Learning (arXiv'25) - Framework and benchmark for lifelong agent learning.
• Darwin Godel Machine (arXiv'25) - Agents that rewrite their own code through evolutionary pressure.
• EvoAgent: Self-evolving Agent with Continual World Model (arXiv'25) - Continual world model for long-horizon task evolution.
• Absolute Zero: Reinforced Self-play Reasoning with Zero Data (arXiv'25) - Self-play reasoning without any training data.
• AutoAgent: Evolving Cognition and Elastic Memory Orchestration (arXiv'26) - Self-evolving framework with evolving cognition and elastic memory.
• Group-Evolving Agents (arXiv'26) - Agent groups as evolutionary units with experience sharing. 71.0% on SWE-bench Verified.
• Agent0: Unleashing Self-Evolving Agents from Zero Data (arXiv'25) - Curriculum and executor competition for self-improvement.
• SEMAG: Self-Evolutionary Multi-Agent Code Generation (arXiv'26) - Self-evolutionary agents that auto-upgrade backbone models. 52.6% on CodeContests.
• SAGE: Multi-Agent Self-Evolution for LLM Reasoning (arXiv'26) - Four co-evolving agents from shared LLM backbone.

Memory Optimization

• Agentic Memory: Unified Long-Term and Short-Term Memory Management (arXiv'26) - Memory operations as tool-based actions with progressive RL training via GRPO.
• MEMORA: Harmonic Memory Representation (arXiv'26) - Balances abstraction and specificity. SOTA on LoCoMo and LongMemEval.
• Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory (arXiv'25) - Production architecture. 26% improvement on LOCOMO, 91% latency reduction.
• TeleMem: Long-Term and Multimodal Memory for Agentic AI (arXiv'26) - 19% higher accuracy, 43% fewer tokens, 2.1x speedup over Mem0.
• A-MEM: Agentic Memory for LLM Agents (arXiv'25) - Self-organizing memory with autonomous management.
• Agent Workflow Memory (ICML'24) - Memory tied to agent workflow patterns.
• MemoryBank: Enhancing Large Language Models with Long-Term Memory (AAAI'24) - Structured long-term memory for LLMs.
• Compress to Impress (ICLR'25) - Compression-based memory for extended dialogues.

Prompt and Behaviour Evolution

• ARTEMIS: Evolutionary Optimization for LLM Agent Configurations (arXiv'25) - Semantically-aware genetic operators for joint agent config optimization. 13.6% on competitive programming.
• E-SPL: Unifying Evolutionary Prompt Search and Reinforcement Learning (arXiv'26) - Joint RL weight updates with genetic operators for system prompt evolution.
• EvoPrompt: Connecting LLMs with Evolutionary Algorithms (ICLR'24) - Evolutionary algorithms for prompt optimization.
• Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution (ICML'24) - Prompts that evolve themselves recursively.
• Large Language Models as Optimizers (OPRO) (ICLR'24) - Using LLMs to optimize their own prompts.
• TextGrad: Automatic Differentiation via Text (Nature'25) - Gradient-like optimization through text feedback.

Tool and Code Evolution

• AlphaEvolve (Google'25) - LLM-driven evolutionary code improvement.
• Learning Evolving Tools for Large Language Models (ICLR'25) - Tools that co-evolve with agent capabilities.
• CREATOR: Tool Creation for Disentangling Abstract and Concrete Reasoning (EMNLP'23) - Agents that create their own tools.
• ToolRL: Reward is All Tool Learning Needs (arXiv'25) - Reinforcement learning for tool use optimization.

Reasoning and Planning

• Reflexion: Language Agents with Verbal Reinforcement Learning (NeurIPS'23) - Agents that learn from mistakes through self-reflection.
• ReflAct: World-Grounded Decision Making via Goal-State Reflection (arXiv'25) - Goal-state reflection improving strategic reliability by 27.7% over ReAct.

Safety, Red-Teaming, and Alignment

• AgenticRed: Optimizing Agentic Systems for Automated Red-teaming (arXiv'26) - Evolutionary red-teaming workflow design. 96% attack success on Llama-2-7B.
• Agent vs. Agent: Automated Red-Teaming for Custom Agentic Workflows (EMNLP'25) - AgentHarm-Gen for adversarial task generation. 162% increase in attack success rate.
• AGENTSAFE: Unified Framework for Ethical Assurance and Governance (arXiv'25) - Design, runtime, and audit controls covering the agentic loop.
• OpenGuardrails: Context-Aware AI Guardrails Platform (arXiv'25) - Context-aware safety detection and model-manipulation prevention.

Embodied AI and Robotics

• RACAS: Controlling Diverse Robots With a Single Agentic System (arXiv'26) - Single architecture controlling ground robots, robotic limbs, and underwater vehicles via natural language.
• RoboClaw: Scalable Long-Horizon Robotic Tasks (arXiv'26) - VLM-driven framework with 25% improvement on long-horizon tasks and 53.7% less human time.
• MEM: Multi-Scale Embodied Memory for Vision Language Action Models (arXiv'26) - Mixed-modal memory for tasks spanning up to fifteen minutes.

Benchmarks and Evaluation

• SWE-bench (ICLR'24) - Can agents resolve real-world GitHub issues?
• AgentBench (ICLR'24) - Multi-dimensional evaluation of LLMs as agents.
• WebArena (ICLR'24) - Realistic web environment for autonomous agents.
• OSWorld (NeurIPS'24) - Open-ended tasks in real computer environments.
• GAIA (ICLR'23) - General AI assistant capabilities benchmark.
• EvoClaw (arXiv'26) - Evaluating agents on continuous software evolution.
• LoCoMo (arXiv'25) - Long-context memory benchmark for agent memory systems.

Community and Knowledge

• Awesome-Self-Evolving-Agents - A comprehensive survey of self-evolving AI agents. Covers single-agent optimization, multi-agent optimization, and domain-specific approaches. by @EvoAgentX (2,037 stars)

Footnotes

Maintained by EvoMap. See contributing guidelines for how to submit a project or paper.

Also check out Awesome Agent Swarm for multi-agent orchestration, swarm intelligence, and collaborative agent systems.