The comprehensive, up-to-date directory of the AI agent ecosystem — frameworks and SDKs for building agents, hosted agent platforms, autonomous coding agents, computer-use/GUI agents, multi-agent orchestration, and the infrastructure underneath it all: protocols, evals, memory, observability, marketplaces, security, and sandboxes.
This list tracks every notable product in the space as of September 2026. A machine-readable copy lives in data/agents.json.
Scope: This is the broad ecosystem list — the starting point for anything agent-related. It covers single-agent frameworks and SDKs, hosted agent development platforms, autonomous coding agents, and computer-use/GUI agents, alongside the supporting layers (protocols, evals, memory, observability, marketplaces, security, sandboxes). Its Multi-Agent Orchestration section overlaps with two sibling lists: when you want depth on coordination itself rather than ecosystem breadth, open awesome-AI-agent-orchestration (the coordination mechanisms: graphs, handoffs, protocols, memory, durable execution) or awesome-multi-agents-workflow (multi-agent teams as working systems: crews, supervisor teams, SaaS/cloud platforms, workflow infrastructure).
Status tags: (beta) = public beta/preview · (legacy) = superseded but still around · (archived) = repo archived · (deprecated) = vendor-retired · (sunsetting) = being wound down — avoid for new work · (maintenance) = maintenance mode only — avoid for greenfield · acquired by X = product continues under new owner
- Agent Frameworks & SDKs
- Agent Development Platforms
- Autonomous Coding Agents
- Computer-Use / GUI Agents
- Multi-Agent Orchestration
- Communication Protocols
- Evals & Benchmarks
- Agent Memory
- Agent Observability
- Agent Marketplaces
- Agent Security & Guardrails
- Agent Sandboxes & Tool Platforms
- Guides
- Status Changes
- Related Lists
- Contributing
- License
Libraries and SDKs for building agents — single-agent loops, tool use, and multi-agent coordination.
- LangChain
OSS— Foundational LLM app framework: chains, prompts, tools, retrievers. - LangGraph
OSS— Graph-based, stateful agent orchestration with checkpointing; supports supervisor, swarm, and hierarchical multi-agent patterns. - LlamaIndex
OSS— Data framework for RAG-centric agentic systems; event-driven Workflows. - Haystack
OSS— Production NLP pipeline framework from deepset for enterprise document-processing agents. - Semantic Kernel
OSS— Microsoft's enterprise AI orchestrator (C#, Python, Java). - Microsoft Agent Framework
OSS— Converged AutoGen + Semantic Kernel agent stack (Python + .NET). - AutoGen
OSS(legacy)— Microsoft's conversational multi-agent framework; pioneered multi-agent group chat — continued by the community as AG2. - AG2
OSS— Community continuation of AutoGen by its original creators; open governance with MCP and A2A support. - CrewAI
OSS— Role-based multi-agent framework; agents, tasks, and crews map to real teams. - OpenAI Agents SDK
OSS— Lightweight, handoff-centric agent framework from OpenAI with tracing and guardrails. - Google ADK
OSS— Google's multi-agent framework with A2A built in; hierarchical agent systems pair with Vertex AI. - Amazon Strands Agents
OSS— AWS-native composable agent framework. - Claude Agent SDK — Anthropic's SDK for building Claude-powered agents; MCP-native with extended thinking (Python and TypeScript SDKs).
- smolagents
OSS— Hugging Face's minimal, code-first agent library. - Pydantic-AI
OSS— Type-safe agent framework from the Pydantic team with validated structured outputs. - Agno
OSS— High-performance agent framework with multimodal support; Teams primitive and AgentOS runtime. - Mastra
OSS— TypeScript-first agent framework with built-in memory, evals, RAG, and MCP. - Vercel AI SDK
OSS— Streaming-first agent SDK for web apps and agent UIs. - MetaGPT
OSS— Multi-agent software-company simulation with SOPs and typed artifact pub/sub between roles. - CAMEL-AI
OSS— Communicative-agent framework for role-play and multi-agent research. - DSPy
OSS— Declarative programming of LM pipelines with automatic prompt optimization (Stanford NLP). - AutoGPT
OSS— The 2023 autonomous-agent pioneer; general-purpose agents that plan and execute tasks.
Hosted platforms for building, running, and managing agents — low-code to fully managed runtimes.
- Vertex AI Agent Builder — Google Cloud's managed agent development platform; framework-agnostic runtime with RAG Engine and Google Search grounding.
- ChatGPT Agent — OpenAI's agentic product for browser automation and long multi-step tasks; absorbed Operator in July 2025.
- Claude Managed Agents
(beta)— Anthropic's hosted agent execution environment; stateful sessions with sandboxing, no own infra needed. - Amazon Bedrock Agents / AgentCore — AWS managed agent service; framework-agnostic, multi-model, with Bedrock Guardrails.
- Azure AI Foundry — Full-stack AI platform with managed agent runtime for Microsoft-centric shops.
- Microsoft Copilot Studio — Low-code agent builder for Microsoft 365, Dynamics, and Power Platform.
- Salesforce Agentforce — CRM-native autonomous agents on live Salesforce data; Atlas Reasoning Engine.
- Relevance AI — No-code multi-agent "workforce" builder with a visual canvas; strong in sales/GTM workflows.
- Lindy — Template-first no-code AI employees for email, CRM, and calendar; "Agent Swarms" for parallel tasks.
- Voiceflow — Omnichannel chat and voice agent design platform; one agent runs across web, app, WhatsApp, SMS, and voice.
- Botpress — Long-standing hosted agent creation and deployment platform.
- Dify
OSS— Open-source LLMOps platform with visual agent builder and broad plugin marketplace. - n8n
OSS— Workflow automation with native AI-agent nodes and MCP support; self-hostable. - Activepieces
OSS— Open-source AI workflow automation; MCP-server-first automation for agents. - Agno AgentOS — Hosted and self-hosted agent operating system behind the Agno framework.
- LiveKit Agents
OSS— Real-time voice agent framework on WebRTC; open source with plugin architecture and MCP support. - Zapier Agents — Custom AI teammates inside Zapier automations, grounded in company knowledge.
- Flowise
OSS(sunsetting)— Visual LangChain/LlamaIndex agent builder; sunsetting — do not build on it.
Agents that write, edit, test, and ship code — terminal CLIs, IDE extensions, and cloud engineers.
- Claude Code — Anthropic's terminal-first agentic coding tool: subagents, hooks, plan mode, MCP.
- Codex CLI
OSS— OpenAI's open-source terminal coding agent; headlesscodex execand MCP server mode. - Cursor — AI-native IDE with background agents; the IDE agent to beat.
- GitHub Copilot — Agent mode and completion across VS Code and JetBrains; BYOK for enterprise.
- Windsurf — Agentic IDE with "Cascade" flow and worktrees.
- Aider
OSS— Open-source CLI pair-programmer with strong git integration; BYO LLM. - Cline
OSS— Open-source autonomous coding agent for VS Code; transparent consumption, headless SDK. - Continue
OSS— Open-source, model-agnostic IDE extension agent; Mission Control UI and cloud agents. - Gemini CLI
OSS— Google's open-source terminal agent with LSP integration. - Devin — Cognition's autonomous cloud software engineer; CLI and API.
- Google Jules — Asynchronous autonomous coding agent; GitHub-integrated, runs in cloud sandboxes.
- OpenHands
OSS— Open-source "developer control center"; SDK, CLI, GUI, and Agent Canvas. - SWE-agent
OSS— Research-grade automated GitHub-issue solver (Princeton); the academic reference coding agent. - opencode
OSS— Open-source terminal AI coding agent (TUI) with plugin support and headless JSON mode. - Goose
OSS— Block's extensible local agent; recipes, MCP-native, Agent Client Protocol support. - Roo Code
OSS— Autonomous agent extension in the Cline lineage; multi-mode with cloud agents. - Kilo Code
OSS— Agentic coding extension for VS Code and JetBrains; multi-mode context and Memory Bank. - Crush
OSS— Charmbracelet's TUI coding agent (Go) with multi-model configs. - Amp — Subagent-based coding agent with issue integration; spun off from Sourcegraph as independent Amp Inc. (Dec 2025).
- OpenClaw
OSS— Self-hosted, always-on assistant harness; chat-reachable and proactive, orchestrates CLI agents.
Agents that operate computers the way humans do — clicking, typing, and seeing screens.
- Anthropic Computer Use
(beta)— Screenshot→coordinate→action API tool loop; Docker Linux reference environment. - Claude in Chrome — Browser extension that drives the user's installed Chrome.
- OpenAI CUA — Computer-Using Agent model via the Responses API
computertool; originally powered Operator. - ChatGPT Agent — OpenAI's agentic product for browser automation and long multi-step tasks; absorbed Operator in July 2025.
- OpenAI Operator
(deprecated)— Standalone browser-automation product (Jan 2025); merged into ChatGPT Agent on July 17, 2025. - Browser Use
OSS— Widely-used open-source browser-agent library (Playwright); CLI 2.0 uses the Chrome DevTools Protocol. - Skyvern
OSS— Enterprise browser-RPA replacement; planner-agent-validator architecture. - Agent S3 (Simular)
OSS— Full-desktop (macOS/Linux/Windows) and Android GUI agent. - OmniParser v2
OSS— Microsoft's vision GUI-grounding module that turns any LLM into a GUI agent; pairs with OmniTool. - AIHawk
OSS— Job-application automation agent; Firefox-based community project. - OSWorld
OSS— Real-computer OS task benchmark/environment across Ubuntu, Windows, and macOS; the standard computer-use benchmark. - AndroidWorld
OSS— Real Android phone-control benchmark from Google Research.
Frameworks whose primary abstraction is multi-agent coordination. (Single-agent frameworks with multi-agent modes — LangGraph, AG2, Google ADK, CrewAI, MetaGPT — also appear above.)
- OpenAI Swarm
OSS(archived)— Lightweight educational multi-agent handoff pattern; archived reference implementation — superseded by the OpenAI Agents SDK. - DeerFlow
OSS— ByteDance's long-horizon "SuperAgent" harness with sub-agents and sandboxes. - OpenAgents
OSS— Protocol-based multi-agent communication and orchestration platform. - ChatDev
OSS— Communicative "virtual software company"; classic chat-chain pipeline. - AgentScope
OSS— Alibaba's production-ready multi-agent framework with visual drag-and-drop tooling. - Agency Swarm
OSS— Swarms of agents with distinct roles; type-safe tools with auto-validation. - Kitaru
OSS— ZenML's durable execution framework: checkpoints, replay, and stateful flows for long-running agent systems.
The standards that let agents talk to tools, to each other, and to frontends.
- MCP (Model Context Protocol)
OSS— Standard for connecting agents to tools, data, prompts, and resources; donated to the Linux Foundation (Agentic AI Foundation) in December 2025. - A2A (Agent2Agent)
OSS— Open protocol for discovery, delegation, and artifacts between opaque agents; Linux Foundation, v1.0.0 spec. - ACP (Agent Client Protocol)
OSS— Connect any editor to any coding agent (Zed); unrelated to IBM's ACP/A2A despite the acronym. - AG-UI
OSS— Stream agent state, events, and human interaction into frontends; created by CopilotKit. - ANP (Agent Network Protocol)
OSS— Decentralized agent identity, discovery, and encrypted collaboration (did:wba); W3C community group. - AP2 (Agent Payments Protocol) — Scoped, revocable payment authorization for agentic commerce; Google/Stripe-backed mandate protocol.
- AGENTS.md
OSS— Portable repo-guidance convention for coding agents; read by OpenCode, Crush, Cline, Cursor, and others.
How the field measures agents — task success, tool-call accuracy, and real-computer benchmarks.
- AgentBench — Eight-environment holistic agent evaluation (ICLR 2024).
- GAIA — Real-world questions requiring web search, tool use, and reasoning; difficulty tiers.
- τ-bench — Customer-support agent eval with realistic tools and a user simulator (Sierra).
- SWE-bench — Resolve real GitHub issues across Python repos; the code-agent gold standard, with a cleaned Verified subset.
- WebArena — Realistic web tasks across e-commerce, GitLab, and social sites; VisualWebArena adds visual reasoning.
- OSWorld — Real-computer OS task benchmark/environment; the standard computer-use benchmark.
- BFCL — Berkeley Function Calling Leaderboard; AST-matched tool-call accuracy at scale.
- ToolBench — Tool use across thousands of real APIs (ToolLLM).
- ClawBench — Everyday online tasks for AI agents; a 2026 consumer-task benchmark.
How agents remember — short-term context, long-term knowledge, and self-editing memory.
- mem0
OSS— Self-improving memory layer for agents and assistants. - Zep — Managed memory on a temporal knowledge graph; community edition deprecated — enterprise continues on the Graphiti engine.
- Graphiti
OSS— Real-time bi-temporal knowledge-graph engine (powers Zep); hybrid semantic + keyword + graph retrieval. - Letta
OSS— Stateful agent platform (ex-MemGPT): self-editing memory blocks; Git-backed context. - Cognee
OSS— Open-source memory engine; self-hosted knowledge graph via an ECL pipeline.
Tracing, debugging, and evaluating agents in development and production.
- LangSmith — Tracing, evals, and prompt management for LangChain/LangGraph.
- Langfuse
OSS— Open-source LLM/agent tracing, evals, and prompt management. 🔀 Acquired by ClickHouse (Jan 2026). - Arize Phoenix
OSS— OTel-native tracing with RAG/agent evaluation; commercial Arize AX on top. - Helicone
(maintenance)— Proxy-based LLM observability; in maintenance mode since March 2026 — avoid for greenfield.
Where agents and agent capabilities get discovered, bought, and distributed.
- AWS Marketplace — AI Agents & Tools — Buy and deploy agents; Bedrock AgentCore-deployable with PAYG, contracts, and private offers.
- Salesforce AgentExchange — Partner agent marketplace for the Salesforce ecosystem.
- Microsoft Copilot Agent Store — Enterprise procurement for Copilot agents inside Microsoft 365.
- Anthropic Agent Skills — Open SKILL.md capability-pack spec and ecosystem.
Keeping agents safe: prompt-injection defense, AI firewalls, model scanning, and programmable guardrails.
- Lakera Guard — Real-time prompt-injection and data-leak protection. 🔀 Acquired by Check Point (Sept 2025).
- Robust Intelligence — AI firewall and automated red teaming. 🔀 Acquired by Cisco.
- Protect AI — Model scanning and ML supply-chain security. 🔀 Acquired by Palo Alto Networks.
- NVIDIA NeMo Guardrails
OSS— Programmable, open-source guardrails toolkit (Colang).
Isolated execution for agent code, plus the tool-calling platforms that let agents act in the world.
- E2B
OSS— Cloud sandboxes for agent-generated code; desktop (computer-use) sandboxes. - Arcade
OSS— Tool-calling platform with per-user authorization; agents act as the user, not a bot. - Composio
OSS— Managed auth and app integrations as agent tools; skips OAuth plumbing for Gmail, Slack, Jira. - Firecracker
OSS— AWS microVMs; the isolation tech underneath most multi-tenant agent execution platforms.
- Choosing an agent framework — how to pick between LangGraph, CrewAI, ADK, the OpenAI Agents SDK, and the rest.
- Agent protocols — MCP vs A2A vs ACP vs AG-UI vs ANP vs AP2, and when each matters.
- Evaluation and observability — benchmarks, tracing, and what to measure before production.
- Status changes — acquisitions, shutdowns, deprecations, and maintenance-mode entries.
- Glossary — the terms this ecosystem can't stop inventing.
Notable recent churn — see docs/status-changes.md for details:
- OpenAI Operator merged into ChatGPT Agent (Jul 2025); OpenAI Swarm archived (superseded by the OpenAI Agents SDK).
- AutoGen is legacy — the community continues it as AG2; Flowise is sunsetting.
- MCP donated to the Linux Foundation (Dec 2025); IBM's ACP merged into A2A (Aug 2025).
- Langfuse acquired by ClickHouse (Jan 2026); Helicone in maintenance mode (Mar 2026); Zep Community Edition deprecated.
- Lakera Guard → Check Point; Robust Intelligence → Cisco; Protect AI → Palo Alto Networks.
- awesome-ai-sandboxes — the deep dive on AI sandboxes and coding-agent sandboxes (this list's Agent Sandboxes section is the short version).
- awesome-microVM — the microVM ecosystem underneath most agent sandboxes.
- awesome-jev — TypeSafe Jev / System One: fast, typed AI decisions with calibrated confidence.
Contributions welcome! Please read CONTRIBUTING.md first — entries need a verified official URL, a one-line description, and no marketing metrics.
This repository is released under the MIT License. Copyright (c) 2026 dakotac1994.