AI agents — every term explained
MCP, tool use, ReAct, agentic loops, planners, and multi-agent systems. Every term you need to build, deploy, or reason about agentic AI in 2026.
AI agents are the fastest-moving area of applied AI in 2026. An agent is an LLM that can take actions in the world — call APIs, browse the web, edit files, run code, delegate to other agents — through tool use. The Model Context Protocol (MCP), introduced by Anthropic and now adopted across major providers, has standardized how agents connect to tools.
The 50 terms in this category are grouped into four sub-topics: core agent concepts (what makes something 'agentic'), the tools and protocols they use (MCP, function calling, computer use), planning and reasoning patterns (ReAct, tree-of-thoughts, planner-executor), and the frameworks people build agents with. For deep coverage of MCP servers and Skills, see our sister site Skills and Servers.
The most important terms in AI Agents
Start here if you're new. These entries explain the foundational vocabulary in depth.
AI Agent
An LLM system that can take actions — call tools, browse the web, edit files, delegate — rather than just producing text. Distinguished from workflow AI by autonomous decision-making about what to do next.
Read the full entry →ProtocolsModel Context Protocol (MCP)
An open standard introduced by Anthropic that lets AI applications connect to external data sources and tools through a uniform interface — the "USB-C for AI." Now supported across major providers.
Read the full entry →ActionsTool Use
The ability for an LLM to invoke external functions — APIs, code execution, file operations. Foundation of agentic behavior, implemented via function calling or MCP.
Read the full entry →PatternsAgentic Loop
The core cycle: observe → decide → act → observe result. Repeats until the task is complete or a stop condition is met.
Read the full entry →PatternsReAct Pattern
Reasoning + Acting — the pattern where the model thinks step-by-step, then takes an action, then observes the result, then thinks again. Introduced in a 2022 paper and still fundamental.
Read the full entry →ActionsComputer Use
An agent capability where the model can control a computer directly — clicking, typing, taking screenshots. Anthropic's Claude and OpenAI's Operator popularized production usage.
Read the full entry →All AI Agents terms, organized
Grouped into sub-topics so you can find neighbors and prerequisites, not just alphabetical entries.
Core agent concepts
Tools & protocols
Patterns & reasoning
Frameworks & tools
Related term categories
These categories connect naturally to AI Agents — many terms cross-reference between them.
Go deeper on AI Agents
📖 Related concept tutorials
Long-form guides that walk through how these concepts actually work.
- How Agentic AI Works
The agentic loop, tools, planning explained.
- How MCP Works
Servers, clients, transports, resources.
- Building a Coding Agent
End-to-end walkthrough with real code.
🎯 Related models & tools
Real products and models where you'll encounter these terms.
- Claude Opus 4.8
Industry-leading agentic capabilities.
- Claude Code
Anthropic's agentic coding tool.
- MCP vs Function Calling
When to use each.
Questions about AI Agents
A chatbot responds to messages. An agent takes actions — calls APIs, edits files, browses the web, delegates to other agents. Every agent is a chatbot under the hood, but not every chatbot is an agent.
Not replacing — layered on top. Function calling is the low-level LLM capability. MCP is a protocol for servers to expose tools that clients (using function calling under the hood) can consume. See MCP vs Function Calling.
No. Simple agents can be built with just the provider SDK (Anthropic, OpenAI) and function calling. Frameworks help with orchestration, memory, and complex flows. Start simple; adopt a framework when the complexity actually appears.
For narrow, well-scoped tasks: yes. For open-ended autonomous behavior: still risky. The current sweet spot is agents with clear boundaries — coding agents, research agents, customer-support triage — where the action space is constrained.
Task success rate on a defined benchmark (SWE-bench for coding, WebArena for browsing) plus operational metrics — tokens used, tools called, time to completion, error recovery rate. See our evaluation category.