Taxonomy — 15 categories · 112 systems
Types of AI agents
agentlist.io groups agents into fifteen categories. Developer-facing: AI IDEs (forked editors with agents built in), IDE extensions, CLI agents, cloud agents (ticket-in, PR-out), platforms, model routers, bridges (protocols connecting agents to tools), agent communication (protocols and frameworks for agents talking to agents — A2A, CrewAI), and agent services (the substrate agents can't self-provision — payments, identity, inboxes, sandboxes). Always-on and role-based: personal agents (self-hosted assistants in your chat apps), digital employees (agents hired for a business role — SDR, support rep, legal associate), ops agents (AI SREs and SOC analysts holding the pager), computer-use agents (driving a GUI instead of an API), research agents (deep-research report generators), and voice agents (answering and making phone calls).
Taxonomy from seed 2026-09-25. Each category page defines the term, explains who it's for, lists tradeoffs, and links every cataloged system in it.
AI IDEs /ai-ide
An AI IDE is a forked or purpose-built code editor whose primary interface is an AI agent: repo-aware chat, multi-file edits, and agent modes are first-class, not a plugin. Most are VS Code forks; some are native editors that add agent panels.
Cursor · Windsurf / Devin Desktop · Kiro · Google Antigravity · Zed · Trae
IDE extensions & copilots /ide-extension
IDE extensions add completion, chat, and agent modes to an existing editor — VS Code, JetBrains, Neovim — through the editor's plugin system. They range from vendor copilots to open-source, bring-your-own-key agents.
GitHub Copilot · Cline · Amazon Q Developer · Gemini Code Assist · Sourcegraph Cody · Augment Code · Continue · Roo Code · Kilo Code · Junie
CLI & terminal agents /cli-agent
CLI agents run in the shell against a local checkout. They read the repository, plan changes, edit files, execute commands and tests, and often commit — usually with tool protocols like MCP for extension.
Claude Code · Codex CLI · Aider · Amp · OpenCode · OpenHands · Goose · Gemini CLI · Warp · Muse Code · Grok Build · Kimi Code · Qwen Code · Mistral Vibe · Crush
Cloud & autonomous agents /cloud-agent
Cloud agents run in a vendor-hosted sandbox rather than on your machine. You hand them a task — an issue, a ticket, a prompt — and they return a pull request or a deployed change for human review.
Devin · Factory Droids · Jules · Replit Agent · Roomote
Platforms & harness multiplexers /platform
Platforms sit above individual agents: multi-agent IDEs that run Claude Code, Codex, and Cursor side by side in Git worktrees, or managed APIs that deploy coding agents across harnesses with sandboxes and credentials.
Cursor Origin
Model routers & gateways /model-router
Model routers expose a single, usually OpenAI-compatible API and route requests across providers with fallbacks, spend controls, and usage billing. Most BYOK coding agents can point at one.
OpenRouter · LiteLLM
Bridges, brokers & protocols /bridge
Bridges attach agents to tools, data, and channels — tool protocols (MCP), memory servers, and transport bindings that expose an agent through chat or voice. For agents talking to agents, see agent communication.
Model Context Protocol · Instinct
Agent communication & collaboration /agent-communication
Bridges connect agents to tools; this category connects agents to each other. It splits into open protocols for cross-vendor discovery and messaging (A2A, AGNTCY, ANP, NANDA) and multi-agent frameworks where a supervisor delegates work to specialists through handoffs (CrewAI, Microsoft Agent Framework, LangGraph, MetaGPT, CAMEL).
OpenScout / Scout · A2A Protocol · AGNTCY · Agent Network Protocol · NANDA · BeeAI · CrewAI · Microsoft Agent Framework · LangGraph · MetaGPT · CAMEL · Agency Swarm · OpenAI Agents SDK
Agent services /agent-services
Agent services are purpose-built primitives whose customer is the agent (or its builder): ways to pay (x402, AP2, AgentCard, Skyfire), ways to authenticate (Keycard, Arcade, Scalekit), places to exist (AgentMail inboxes, E2B sandboxes, Browserbase browsers), and senses (Exa search, Mem0 memory). Where bridges give agents protocols, services give them accounts — metered, billable capabilities the agent can't synthesize.
AgentMail · x402 · Agent Payments Protocol · Skyfire · Payman · Nevermined · AgentCard · Keycard · Arcade · Scalekit · Composio · E2B · Browserbase · Exa · Mem0
Personal & always-on agents /personal-agent
Personal agents are self-hosted assistant runtimes: one gateway or process that holds sessions, memory, tools, and skills, and meets you in the channels you already use — Telegram, WhatsApp, Discord, Slack, Signal, iMessage — alongside CLIs, TUIs, and web dashboards. They run on your hardware or your VPS, keep state on your machine, and delegate to repo-native coding agents when the task is code.
OpenClaw · Hermes Agent · NanoClaw · ZeroClaw · PicoClaw · nanobot
AI employees & digital coworkers /digital-employee
AI employees are agents packaged as a job function — an SDR that prospects, a support rep that resolves tickets, a recruiter that screens candidates. They arrive with an identity, a role, and integrations into the systems that role touches (CRM, helpdesk, ATS, email, phone), and they're sold on seats or outcomes rather than editor features.
Alice · Ava · Piper · Sierra · Decagon · Fin · Harvey · Lindy · Juicebox (PeopleGPT) · Mercor
On-call & ops agents /ops-agent
Ops agents sit inside the incident and security loop: they watch alerts, correlate telemetry, investigate pages, draft root-cause analyses, and propose or apply remediations. AI SREs handle production incidents; AI SOC analysts (Dropzone, Simbian) handle security queues.
Cleric · Resolve · incident.io · Rootly · Bits AI · Dropzone AI · Simbian
Browser & computer-use agents /computer-use
Computer-use agents operate software through its interface — a browser, a desktop, a remote session — by reading the screen or DOM and acting with clicks, keystrokes, and navigation. Some are hosted products that take a goal and return a result (ChatGPT Agent, Manus); others are open frameworks and APIs for building your own (Browser Use, Skyvern, Stagehand, Claude Computer Use).
Browser Use · Skyvern · ChatGPT Agent · Claude Computer Use · Manus · Stagehand
Research & analyst agents /research-agent
Research agents take a question and return a structured, cited report — searching, reading, and synthesizing dozens of sources over minutes-long runs. Consumer 'deep research' modes live inside chat products (ChatGPT, Gemini, Perplexity); enterprise analysts work over licensed or internal corpora (Hebbia, AlphaSense); academic tools ground in the literature (Elicit).
ChatGPT Deep Research · Gemini Deep Research · Perplexity · Genspark · Hebbia (Matrix) · Elicit · AlphaSense
Voice & phone agents /voice-agent
Voice agents handle real-time spoken conversations: inbound reception and support, outbound reminders and qualification, scheduling, intake, and surveys. The category is mostly developer platforms (Vapi, Retell, Bland, Deepgram, ElevenLabs Agents) that orchestrate speech-to-text, a model, and text-to-speech over telephony — plus no-code products (Synthflow) that ship a ready receptionist.
Vapi · Bland AI · Retell AI · Synthflow · ElevenLabs Agents · Deepgram Voice Agent · Grok Voice Agent API
Frequently asked
What are the main types of AI agents?
agentlist.io groups agents into fifteen categories. Developer-facing: AI IDEs (forked editors with agents built in), IDE extensions, CLI agents, cloud agents (ticket-in, PR-out), platforms, model routers, bridges (protocols connecting agents to tools), agent communication (protocols and frameworks for agents talking to agents — A2A, CrewAI), and agent services (the substrate agents can't self-provision — payments, identity, inboxes, sandboxes). Always-on and role-based: personal agents (self-hosted assistants in your chat apps), digital employees (agents hired for a business role — SDR, support rep, legal associate), ops agents (AI SREs and SOC analysts holding the pager), computer-use agents (driving a GUI instead of an API), research agents (deep-research report generators), and voice agents (answering and making phone calls).
What is the difference between an AI IDE and an IDE extension?
An AI IDE is a whole editor — usually a VS Code fork — redesigned around an agent, so the vendor controls the entire UI. An IDE extension installs into the editor you already use (VS Code, JetBrains, Neovim) and adds agent features without replacing it. Extensions are lower-commitment; AI IDEs can integrate more deeply.
What is the difference between a CLI agent and a cloud agent?
A CLI agent runs on your machine, in your terminal, against your local checkout — you watch and steer it in real time. A cloud agent runs on the vendor's infrastructure in its own sandbox, typically takes a ticket or prompt, and returns a pull request. CLI agents favour interactive work; cloud agents favour delegating whole tasks.
Do I need a model router or a bridge to use a coding agent?
No. Most agents work out of the box with their default model provider. A model router (like OpenRouter or LiteLLM) is useful when you want to swap providers, control cost, or use one API key across many models. A bridge (like the Model Context Protocol) is useful when you want an agent to reach tools, databases, or internal services in a standard way.