AI-agent glossary
Definitions of 28 core AI-agent terms — MCP, A2A, llms.txt, RAG, HTTP 402, x402, guardrails, and more — each linked to the ChangeGamer resource that covers it in depth.
- A2A (Agent2Agent protocol)
- An open protocol for communication between autonomous agents: agents publish capability descriptions (agent cards) and exchange tasks and results with each other, complementing protocols like MCP that connect an agent to tools.
- Agent evals
- Systematic evaluation of agent behavior — task success, trajectory quality, safety — using scored test suites rather than ad-hoc manual inspection, so regressions are caught when models or prompts change.
- Agent observability
- Tracing and monitoring of agent runs — model calls, tool invocations, tokens, latencies, errors — so operators can debug failures and track cost and quality in production.
- AGENTS.md
- A Markdown file convention placed in a code repository to give coding agents project-specific instructions — build commands, conventions, and rules — that they read before making changes.
- AI agent
- A software system that uses a large language model to decide and act toward a goal — calling tools, reading and writing data, and iterating on results — rather than only generating text for a human to read.
- Chunking
- Splitting documents into retrieval-sized pieces before embedding and indexing them for RAG. Chunk size and boundaries strongly affect retrieval quality.
- Computer use
- An agent capability where the model operates a real user interface — clicking, typing, and reading screens in a browser or desktop — instead of calling structured APIs.
- Durable execution
- Running long-lived workflows so that state survives crashes and restarts: each step is persisted and resumable, which matters for agents whose tasks span minutes, hours, or external approvals.
- Embedding
- A numeric vector representation of text (or other data) in which semantic similarity becomes geometric closeness, enabling similarity search, clustering, and retrieval.
- Fine-tuning
- Continuing a model's training on task- or domain-specific examples to change its default behavior — as opposed to supplying knowledge at query time via retrieval or context.
- Guardrails
- Technical controls around an agent — input/output filtering, policy checks, action allowlists, human-approval gates — that constrain what it can do regardless of what the model generates.
- HTTP 402 (Payment Required)
- The HTTP status code reserved for payments. Agent-first sites use it to gate premium content machine-readably: the response body carries the price, a payment URL, and retry instructions, so an agent can pay and retry without a human.
- Large language model (LLM)
- A neural network trained on large amounts of text to predict and generate language. In an agent system it is the reasoning and generation component: the model that interprets instructions, reasons about a task, and decides what to say or which tool to call next.
- llms.txt
- A plain-text/Markdown file convention at a site's root that gives language models and agents a curated, machine-friendly index of the site's content — analogous in spirit to robots.txt, but for content discovery rather than crawl permission.
- MCP (Model Context Protocol)
- An open protocol that standardizes how AI applications connect to external tools, data sources, and prompts. Servers expose capabilities (tools, resources, prompts); clients such as agents and chat apps consume them over defined transports.
- MCP server
- A program implementing the server side of the Model Context Protocol: it advertises tools, resources, and prompts that any MCP client can discover and invoke.
- Multi-agent orchestration
- Coordinating several specialized agents on one task — via supervisor hierarchies, pipelines, or peer protocols — including how work is delegated, results are merged, and failures are contained.
- OAuth
- An open authorization framework in which a user or system grants a third-party application limited, scoped access to protected resources by issuing access tokens, without sharing the underlying credentials. Remote MCP servers commonly use it to authenticate and authorize connecting clients.
- Prompt caching
- Provider-side reuse of computation for repeated prompt prefixes, cutting cost and latency for agents that resend large, stable system prompts and tool definitions on every call.
- Prompt injection
- An attack where instructions hidden in data an agent processes (a web page, email, or document) manipulate the agent into ignoring its real instructions or misusing its tools. A core security concern for tool-using agents.
- RAG (Retrieval-Augmented Generation)
- An architecture where a model's answer is grounded in documents retrieved at query time — typically via embedding search over an indexed corpus — instead of relying only on what the model memorized in training.
- Reranking
- A second-stage retrieval step where a stronger (usually cross-encoder) model re-orders an initial candidate list so the most relevant documents reach the model's context first.
- robots.txt
- A plain-text file at a website's root, defined by the Robots Exclusion Protocol, that tells web crawlers — including AI crawlers and agents — which parts of the site they may or may not access.
- RSL (Really Simple Licensing)
- A machine-readable licensing vocabulary that lets publishers declare, in an XML file, the terms under which automated systems may use their content — including AI-training permissions and pricing.
- Streaming
- Delivering model output incrementally (typically over server-sent events) as it is generated, so users and downstream agents can act on partial results instead of waiting for the full response.
- Structured outputs
- Model-API features that constrain generation to a schema (typically JSON Schema), so downstream code can parse model output without brittle prompt-and-pray formatting.
- Tool calling (function calling)
- The mechanism by which a model requests execution of a developer-defined function with structured arguments, receives the result, and continues reasoning — the basic building block of agent behavior.
- x402
- An open standard built on HTTP 402 for fully programmatic payments: the 402 response carries structured payment requirements, the client retries with a signed payment header, and a facilitator verifies and settles the payment on-chain.