All terms

A2A Task: An A2A Task is the core stateful unit of work in the Agent2Agent (A2A) protocol.
ACP (four protocols): ACP refers to four distinct specifications from four groups across three steps of the map.
Agent Card: An Agent Card is a JSON document that an A2A server publishes to describe itself to other agents: its identity (name, description, provider), its service endpoint URL, the A2A capabilities it supports (streaming, push notifications), its authentication schemes, and the skills it offers, each with an id, description, and input/output modes.
Agent skill: An Agent Skill is a portable, self-contained folder of domain instructions, workflow rules, and helper scripts—anchored by a SKILL.md file with YAML frontmatter (name, description)—that an AI coding or research agent discovers at startup and loads on demand when a task matches its description.
Agent2Agent (A2A): Agent2Agent (A2A) is an open protocol—launched by Google in April 2025 and stewarded by the Linux Foundation Agentic AI Foundation (AAIF)—for peer-to-peer communication and task delegation between opaque autonomous agents across organizational boundaries.
Agentic EoR (Employer / Entity of Record): An Agentic EoR (Employer of Record or Entity of Record) is a legal, financial, and cryptographic intermediary that acts as the accountable legal counterparty when an autonomous software agent dispatches work, hires human contractors, enters service contracts, or settles cross-border payments [1].
Agentic Resource Discovery (ARD) manifest: An ARD manifest (/.well-known/ard.json, superseding /.well-known/ai-catalog.json in ARD v0.91) is a JSON document published at an origin's well-known path that inventories every agent-facing resource on that domain in a single fetch: MCP endpoints, SEP-2127 Server Cards, A2A Agent Cards, OpenAPI specifications, /llms.txt, and payment/auth metadata.
Agentic web: The agentic web is the ordinary web plus a layer of conventions that let software agents, rather than people in browsers, discover what a site offers, prove who they are, learn what they may do, read content, invoke capabilities, coordinate with other agents, pay, present results, and be held to account.
Agentmap directive: Agentmap is a line a publisher can add to robots.txt, of the form Agentmap: https://example.com/entries.json, that points a crawler at a source of Agentic Resource Discovery (ARD) entries as an alternative to relying solely on the well-known manifest path.
AI agent: An AI agent (or software agent) is an autonomous or semi-autonomous program that pursues a user-delegated goal by planning across steps, reading external state, and invoking tools or network endpoints without a human driving each click.
AI crawler tokens: AI crawler tokens are the User-agent names AI companies document for use in robots.txt under RFC 9309, split by purpose rather than treated as one identity.
API Catalog (RFC 9727): api-catalog (RFC 9727, published June 2025) is an IANA-registered well-known URI (/.well-known/api-catalog) and HTTP link relation (rel="api-catalog") that returns an RFC 9264 application/linkset+json document listing a publisher's machine-readable APIs, OpenAPI specifications (service-desc), documentation (service-doc), and status endpoints (status).
Audit record: An audit record in the agentic web is a structured, tamper-evident log entry capturing what action an agent took, which tool or endpoint it invoked, which upstream principal or mandate authorized the call, and the cryptographic trace context (traceparent / WIMSE Wit) linking the causal chain across multi-agent hops.
Buyer of Record & Merchant of Record (Agentic Commerce): In agentic commerce, the Buyer of Record (BoR) and Merchant of Record (MoR) define which legal entities bear tax, chargeback, refund, and regulatory responsibility when an AI agent or harness executes a purchase on a user's behalf [1].
Coding agent instruction files: AGENTS.md is an open repository and origin convention (agents.md, stewarded under the Linux Foundation AAIF) for placing machine-targeted instructions—build commands, test runners, architectural invariants, and endpoint conventions—at the root of a git repository or website (/AGENTS.md).
Content usage preference: A content usage preference is a machine-readable declaration by a publisher stating how automated systems may use a resource—for example, permitting search indexing and live user-triggered grounding while prohibiting foundation-model training.
Content-Signal directive: Content-Signal is a machine-readable policy directive introduced by Cloudflare's Content Signals Policy that publishers place in robots.txt or HTTP response headers (Content-Signal: search=yes, ai-input=yes, ai-train=no) to express separate permissions for search indexing (search), live RAG/grounding retrieval (ai-input), and foundation model training (ai-train).
Context compaction: Context compaction is the set of protocol and runtime patterns used to keep an AI agent's working context window bounded across multi-turn tool invocations—preventing attention degradation and redundant token spend.
Dynamic Client Registration & Client ID Metadata: Dynamic Client Registration (DCR, RFC 7591) and OAuth Client ID Metadata Documents allow an AI agent or MCP host to authenticate with a newly discovered authorization server without a human developer manually pre-registering a client_id on that site first.
Elicitation: Elicitation (elicitation/create) is a capability in the Model Context Protocol (MCP) that lets an MCP server pause mid-execution and ask the human user—via the client host—for structured additional information validated against a flat JSON Schema, or direct the user to an external URL for out-of-band confirmation.
Expired Internet-Draft: An Internet-Draft that has not been revised within 185 days.
Facilitator: A facilitator in the x402 protocol is an optional third-party service that verifies and settles blockchain payments on behalf of a resource server, so the server does not need its own blockchain connectivity or payment-verification logic.
Generative UI: Generative UI is an interaction pattern in which an AI agent returns structured, declarative user-interface components—such as interactive forms, data tables, comparison cards, or sandboxed views—rather than plain text or Markdown.
HTTP 402 Payment Required: HTTP 402 Payment Required is a standard HTTP status code (reserved since HTTP/1.1 and defined in RFC 9110 §15.5.3) indicating that a request cannot be fulfilled until the client makes a payment.
HTTP Message Signatures: HTTP Message Signatures (RFC 9421, published February 2024 by the IETF HTTP Working Group) define a standard mechanism for creating, encoding, and verifying digital signatures or MACs over selected components of an HTTP message.
Human Not Present: Human Not Present (HNP) is an execution mode in Google's Agent Payments Protocol (AP2) and FIDO-aligned agentic commerce frameworks in which an autonomous agent completes a purchase when the user is offline or not actively viewing a checkout screen.
Idempotency-Key header: An Idempotency-Key is a unique client-generated token attached to an HTTP POST/PATCH request or agent payment invocation so that if an agent retries after a network timeout, the server recognizes the duplicate and returns the original result rather than executing the action or charging the user twice.
Interrupt: In AG-UI (Agent-User Interaction Protocol) and stateful agent runtimes, an interrupt is a structured lifecycle event where an agent pauses an in-flight execution run to request human confirmation, disambiguation, or credentials (human-in-the-loop), and then resumes the run from its checkpointed state once the user responds.
Know Your Agent: Know Your Agent (KYA) means verifying which agent, and which operator behind it, is transacting before trusting it.
llms.txt: llms.txt is a Markdown index file placed at /llms.txt (with an optional expanded companion at /llms-full.txt) proposed by Jeremy Howard (Answer.AI) in September 2024 to give language models and coding agents a concise, token-efficient map of a website's most important documentation and canonical URLs.
Mandate: A mandate in Google's Agent Payments Protocol (AP2) is a cryptographically signed, verifiable credential that authorises an agent to act on a person's behalf for a specific purchase step.
Markdown twin: A Markdown twin is a clean text/markdown representation of an HTML web page served either at a predictable sibling URL (such as /path/index.md or /page.md) or via HTTP content negotiation when an agent sends Accept: text/markdown.
MCP: The Model Context Protocol (MCP) is an open JSON-RPC 2.0 standard—originated by Anthropic in November 2024 and stewarded under the Linux Foundation Agentic AI Foundation (AAIF)—that connects AI host applications (IDEs, chat desktops, autonomous runtimes) to external tool and context servers.
MCP Prompt: An MCP Prompt is a reusable, parameterized prompt template or workflow slash-command exposed by an MCP server via prompts/list and prompts/get.
MCP Registry: The Official MCP Registry (modelcontextprotocol/registry) is the community-governed catalogue and HTTP API of public Model Context Protocol servers, maintained by the MCP Registry Working Group under the Linux Foundation AAIF.
MCP Resource: An MCP Resource is a URI-addressed piece of read-only context data—such as a file, database record, documentation page, or live telemetry snapshot—that an MCP server exposes via resources/list, resources/templates/list, and resources/read.
MCP Server Card: A Server Card is a proposed static, discoverable document that describes an MCP server: what it is, who publishes it, and how to reach it, without requiring a client to first open an MCP session.
Origin trial: An origin trial is Chrome's mechanism for shipping an experimental web-platform feature to real sites for a limited time before deciding whether to ship it broadly: a developer registers a specific origin, receives a token, and includes that token as a meta tag or HTTP header so the feature activates for visitors to that origin without requiring them to flip browser flags.
Pay per crawl: Pay per crawl is an HTTP monetization model where web publishers charge AI crawlers and retrieval agents per request using HTTP 402 Payment Required headers rather than blocking them outright in robots.txt or serving them for free.
Protected Resource Metadata: OAuth 2.0 Protected Resource Metadata (RFC 9728, published April 2025) is a discovery document served at /.well-known/oauth-protected-resource that tells an OAuth client or AI agent which authorization server(s) protect an API or MCP server, which scopes_supported apply, and how bearer tokens must be presented.
Roots: Roots (roots/list) in the Model Context Protocol (MCP) are URI boundaries—typically file:// workspace directories or repository paths—that an MCP client host advertises to a connected server to define which filesystem trees or resource scopes the server is expected to operate within.
Sampling: Sampling (sampling/createMessage) is a client-exposed capability in the Model Context Protocol (MCP) that allows an MCP server to request an LLM completion from the host application rather than bundling its own model API keys.
Signed agent: A signed agent is an automated HTTP client or AI agent that cryptographically signs its outbound requests using RFC 9421 HTTP Message Signatures and publishes its public verification keys at a discoverable key directory (such as /.well-known/http-message-signatures-directory), allowing origins and CDNs to verify the agent's identity independent of IP address or spoofable User-Agent strings.
SKILL.md: SKILL.md is the entrypoint manifest of an Agent Skill folder (agentskills.io).
Streamable HTTP: Streamable HTTP is the primary remote transport for the Model Context Protocol (MCP), replacing the early dual-endpoint HTTP+SSE transport from the 2024-11-05 specification.
Subagent: A subagent is an isolated child agent invocation spawned by a parent orchestrator agent to execute a bounded subtask in its own context window, returning a synthesized result rather than polluting the parent's trajectory with raw intermediate tool outputs.
Tool: In the Model Context Protocol (MCP), a tool is a named, JSON-Schema-typed function exposed by an MCP server (tools/list and tools/call) that a host application allows a language model to invoke to inspect external state or perform side effects.
Verifiable Intent: Verifiable Intent is a cryptographic trust specification—developed by Mastercard and contributed to the FIDO Alliance Payments Technical Working Group in April 2026—that uses chained SD-JWT (Selective Disclosure JSON Web Token) credentials to prove what a human user authorized an AI agent to purchase.
Web Bot Auth: Web Bot Auth is Cloudflare and IETF-draft profile of RFC 9421 HTTP Message Signatures whereby an AI crawler or agent signs its outbound HTTP requests with an Ed25519 private key and advertises its public key set (JWKS) at /.well-known/http-message-signatures-directory via a Signature-Agent header.
Well-known URI: A well-known URI is a path beginning /.well-known/ at the root of an origin, reserved by RFC 8615 (May 2019, obsoleting RFC 5785) so that applications needing site-wide metadata do not have to negotiate arbitrary paths or risk colliding with a site's existing URL space.
Workload identity: Workload identity is a cryptographic identity issued to a software process—such as a microservice, container, or autonomous AI agent—rather than to a human user.
x402 Payment Payload: An x402 Payment Payload is the base64-encoded JSON credential carried in the HTTP X-PAYMENT (v1) or PAYMENT-SIGNATURE (v2) request header when an agent pays for an HTTP 402 Payment Required resource under the x402 protocol.