AI Glossary

AI Agent AI Agent A software system that perceives its environment, decides its own next step, and calls tools to reach a goal — not a single-turn chat program. Agentic AI Agentic AI An umbrella label for AI systems that emphasise autonomous decision-making and multi-step execution; it names a design orientation, not a model type. Autonomous Agent Autonomous Agent An agent that runs multiple steps to completion without a human confirming each one — usually contrasted with approval-gated agents. Multi-Agent System Multi-Agent System An architecture in which several agents with distinct roles and independent contexts collaborate on one overall task. Subagent Subagent An agent spawned by a parent agent to handle one bounded subtask, with its own context window and tool permissions, returning a result to the parent. Agent Orchestration Agent Orchestration The control layer that decides which agent runs when, how outputs flow between agents, and when the whole process terminates. Agent Team Agent Team A set of agents in fixed, specialised roles collaborating persistently on a shared goal — a stable "team" rather than ad-hoc spawning. Agent Loop Agent Loop The repeating cycle observe state → model decides → execute tool → feed result back; the underlying structure of almost every agent framework. ReAct (Reason + Act) ReAct A prompting paradigm (Yao et al., 2022) where the model alternates "Thought" and "Action" and reads back each "Observation" — the ancestor of the modern agent loop. Planning Planning An agent's ability to break a large goal into an ordered sequence of executable steps, before or during execution. Reflection / Self-Correction Reflection / Self-Correction A mechanism where an agent evaluates its own output, spots problems, and produces a revised version — a common way to lower error rates. Human-in-the-loop Human-in-the-loop A design pattern that inserts human approval at key points in an agent's process — the mainstream safety strategy for enterprise deployments. Autonomy Level Autonomy Level A grading idea borrowed from the L0–L5 levels of driving automation, describing how far an agent runs without human intervention. No agreed standard exists yet. Agent-Computer Interface (ACI) Agent-Computer Interface Interface and tool surfaces designed for agents rather than humans, prioritising parseability and low ambiguity over visual presentation. Still an emerging concept.

MCP (Model Context Protocol) MCP (Model Context Protocol) An open standard from Anthropic (Nov 2024) that lets agents discover and call external tools, data, and prompts in a uniform way — often called "USB-C for AI apps." MCP Server MCP Server A service that exposes tools, resources, and prompt templates over MCP for agents to call — the largest category of project in this directory. MCP Client MCP Client The component, usually inside a host app like Claude Code or Cursor, that connects to an MCP server, negotiates capabilities, and invokes its tools for the model. A2A Protocol (Agent2Agent) A2A Protocol (Agent2Agent) An open protocol from Google (April 2025, later donated to the Linux Foundation) for agents from different vendors to discover each other and delegate tasks. ACP (Agent Communication Protocol) ACP (Agent Communication Protocol) A RESTful open standard for agent interoperability from IBM and BeeAI (May 2025); it has since merged into A2A and is no longer developed separately. AGENTS.md AGENTS.md A convention for a Markdown file at a repo's root that tells coding agents how to build, test, and follow the project's conventions. Introduced by OpenAI in 2025. SKILL.md SKILL.md The file that defines an Agent Skill: YAML frontmatter (`name`, `description`) plus body instructions, loaded progressively by the model on demand. Function / Tool Calling Function / Tool Calling The model's ability to request an external function in a structured format and continue once it has the result — the basis of nearly all agent capability. Tool Manifest Tool Manifest A structured declaration of a tool's name, purpose, and parameter schema (usually JSON Schema) that the model reads to decide how to call it. Capability Discovery Capability Discovery An agent querying at runtime what capabilities a server or peer agent currently offers, instead of hard-coding the capability list in code.

Skill Skill A reusable unit packaging instructions, scripts, and reference material for an agent to load on demand; Claude Skills is the canonical implementation. Workflow Workflow A predefined sequence of steps. Anthropic separates "workflows" (fixed code path, LLM works within nodes) from "agents" (the LLM directs its own process). Playbook Playbook A complete procedure for a class of task, written for an agent to follow — usually with decision branches and handling rules, more complete than a single Skill. Pipeline Pipeline Processing stages chained in a fixed order, each stage's output feeding the next; often used for data-processing or content-generation agent tasks. Hook Hook User code that runs automatically at fixed points in an agent's lifecycle (e.g. before/after a tool call), used to audit, block, or rewrite actions. Slash Command Slash Command A predefined instruction triggered by a leading `/`, letting a user invoke a fixed procedure or prompt template precisely without writing natural language. Task Decomposition Task Decomposition The process of splitting a complex goal into independently executable subtasks; it is the dividing line for whether a long task succeeds. Chain-of-Thought Chain-of-Thought A prompting technique (Wei et al., 2022) that has the model write intermediate reasoning steps before its answer, improving accuracy on multi-step tasks. DAG Workflow DAG Workflow Organising task steps as a directed acyclic graph — nodes are steps, edges are dependencies — allowing parallel branches and joins. Common in orchestration tools. State Machine Agent State Machine Agent Constraining an agent with explicit states and transition rules — more controllable than free reasoning, common in support and process-handling systems. Trigger Trigger The event source that starts one agent run. It can be a user message, a timer firing, a webhook from an external system, or another agent's output. Scheduled Agent Scheduled Agent An agent that runs automatically on a fixed cadence such as cron, with no human trigger — used for monitoring, scraping, digests, and periodic checks.