AI Agent Vocabulary.

An agent's think → call tool → observe → repeat loop until the goal is met. Most modern AI apps are this loop.

System for an agent to store and recall information within or across sessions. Short, long, episodic, semantic flavors.

LangGraph's state-machine model where nodes = agents/tools and edges = handoffs. Makes complex flows visualizable.

A loosely coupled, parallel-running group of many small agents. Ant-colony logic; OpenAI Swarm exemplifies this idea.

Front-layer that routes incoming prompts to the most fitting specialist agent. Optimizes cost and quality.

Full log of an agent run including every LLM call, tool, observation. LangSmith, Braintrust, Helicone surface it.

A workflow chaining planning + tools + verification steps instead of a single LLM call.

Agent doing query → search → evaluate → new query in a loop instead of one search. The basis of modern web research.

When the model decides to invoke a function instead of generating text. Same as "function calling." The core of agentic behavior.

Pushing a model to reason step-by-step to solve harder problems. Often as simple as "think step by step."

Mode where the agent takes all actions without human approval. "Full auto", "--dangerously-bypass". Trust mandatory.

Mode where every risky action prompts the user for approval. Claude Code default; the opposite of YOLO.

Agent driving a browser to navigate, fill forms, and scrape. Built on libraries like browser-use or Playwright MCP.

Agent specialized for software work: Claude Code, Cursor Composer, Aider, Devin, Codex, OpenHands.

Anthropic's TypeScript/Python agent library. First-class tool use + computer use + streaming.