GLOSSARY

The vocabulary of the agent channel

Answer-first definitions for the terms that decide whether AI agents can find, understand, and call an API — from the artifacts (llms.txt, AGENTS.md, MCP) to the measurements behind the Discry Score.

Agent infrastructure

llms.txtllms-txt: llms.txt is a plain-markdown index file served at a website's root path (/llms.txt) that tells AI systems where the pages that matter are, so an agent can route straight to the right documentation instead of crawling HTML.
llms-full.txtllms-full-txt: llms-full.txt is a companion convention to llms.txt: the complete documentation of a site concatenated into a single markdown file served at /llms-full.txt, so an agent can load the entire docs surface in one request.
AGENTS.mdagents-md: AGENTS.md is an open convention for a markdown file, placed at the root of a code repository, that briefs AI coding agents on how to work with the project: setup commands, build and test steps, code conventions, and constraints.
Model Context Protocol (MCP)model-context-protocol: The Model Context Protocol (MCP) is an open protocol that standardizes how AI applications connect to external tools and data sources.
MCP servermcp-server: An MCP server is a program that exposes capabilities to AI applications over the Model Context Protocol through three building blocks: tools (schema-defined functions the model can call), resources (read-only data such as documents or schemas), and prompts (reusable instruction templates).
MCP manifestmcp-manifest: An MCP manifest is a JSON file served at the well-known path /.well-known/mcp.json that advertises a domain's Model Context Protocol server, so an agent can determine from the domain alone that an MCP surface exists and how to reach it.
MCP registrymcp-registry: An MCP registry is a public directory of Model Context Protocol servers that agents and developers search to find an MCP surface for a given service.
Tool use (function calling)tool-use: Tool use, also called function calling, is the mechanism by which a language model invokes external functions: the host application declares tools with typed schemas, the model emits a structured call with arguments, and the host executes it and returns the result into the model's context.
AI agentai-agent: An AI agent is a system in which a language model plans and executes multi-step work toward a goal, using tools — web fetches, code execution, API calls — and feeding each result back into its next decision.
Agentic workflowagentic-workflow: An agentic workflow is a multi-step process in which an AI agent chains model calls, tool invocations, and decisions to complete a job — authenticate, create a customer, attach a payment method, then charge it — carrying state between steps and handling failures along the way.
Context windowcontext-window: A context window is the maximum amount of text, measured in tokens, that a language model can process in a single request — the working memory that must hold the system prompt, the conversation, fetched documents, and tool results all at once.
Tokentoken: A token is the unit of text a language model reads and generates — a word fragment averaging roughly four characters of English for common tokenizers — and the unit in which model usage is priced and context windows are sized.
System promptsystem-prompt: A system prompt is the instruction text an application places at the start of a language model's context to define its role, rules, and available tools before any user input arrives.
GPTBotgptbot: GPTBot is OpenAI's web-crawler user agent, used to collect publicly available content that may contribute to training OpenAI's generative AI foundation models.
ClaudeBotclaudebot: ClaudeBot is Anthropic's web-crawler user agent, used to collect public web content that may contribute to training its Claude models.
Google-Extendedgoogle-extended: Google-Extended is Google's robots.txt control token for AI training: publishers use it to manage whether content Google crawls from their sites may be used to train future Gemini models and for grounding — supplying content to the model at prompt time — in Gemini apps and Vertex AI.
AI crawlerai-crawler: An AI crawler is an automated user agent operated by an AI company to fetch web content for model training, for search grounding, or on demand when a user or agent requests a page.
Headless browsingheadless-browsing: Headless browsing is running a web browser without a visible interface so that software can load pages, execute JavaScript, and read the fully rendered result — the technique tools like Playwright and Puppeteer automate.

API documentation

OpenAPI specopenapi-spec: An OpenAPI specification is a machine-readable description of a REST API — its endpoints, parameters, request and response schemas, and authentication schemes — written in JSON or YAML against the OpenAPI Specification standard.
operationIdoperation-id: operationId is an OpenAPI field that assigns a unique, stable identifier to a single API operation — one HTTP method on one path.
API referenceapi-reference: An API reference is the section of documentation that describes every endpoint in contract-level detail: URL, HTTP method, parameters, request and response schemas, error codes, and examples.
Idempotency keyidempotency-key: An idempotency key is a unique, client-generated value sent with an API request so the server can recognize retries of the same operation and execute it only once.
Rate limitingrate-limiting: Rate limiting restricts how many requests a client may make to an API within a given window, enforced per API key, user, IP address, or endpoint.
Pagination (cursor and offset)pagination: Pagination splits a large result set across multiple API responses.
Webhookwebhook: A webhook is an HTTP callback that inverts the usual request flow: the API provider sends an HTTP POST to a URL the client registers whenever a subscribed event occurs, so the client learns about changes without polling.
API versioningapi-versioning: API versioning is the practice of publishing changes to an API under explicit version identifiers — a URL path segment like /v2/, a request header, or a date-based version — so existing integrations keep working while the interface evolves.
Error response formaterror-response-format: An error response format is the consistent structure an API uses for failed requests: the HTTP status code plus a machine-readable body, typically carrying a stable error code, a human-readable message, and often a pointer to the offending field or a documentation link.
Retry and backoffretry-backoff: Retry with backoff is the client pattern of reattempting a failed API request after a delay that grows with each attempt — usually exponential backoff, doubling the wait each time, with random jitter added so many clients do not retry in synchronized waves.
API authentication (API key, OAuth 2.0, PAT)api-authentication: API authentication is how an API verifies the identity of a caller before serving a request.
Sandbox environmentsandbox-environment: A sandbox environment is a separate instance of an API where developers can make real calls against test data without touching production systems or incurring live charges.
SDKsdk: An SDK (software development kit) is an official client library that wraps an API's HTTP interface in idiomatic code for a specific language, handling authentication, serialization, and often retries and pagination, so integrators call typed methods instead of constructing raw requests.
Developer portaldeveloper-portal: A developer portal is the public website where an API publishes what integrators need: reference documentation, guides, authentication setup, changelogs, and often interactive tools such as API explorers and key management.

Discry methodology

Agent-readinessagent-readiness: Agent-readiness is the degree to which AI agents can find and understand an API from its public documentation surface — the pages, specs, and entry-point files an agent actually fetches.
Discry Scorediscry-score: The Discry Score is a behavioral measurement of an API's agent-readiness, taken from its public documentation surface and always decomposed into discovery and comprehension.
Discoverabilitydiscoverability: Discoverability is whether AI agents can discover and understand an API from its public documentation surface.
Discovery layerdiscovery-layer: The discovery layer is the machine-legible surface of an API: the entry-point files and structured signals — llms.txt, AGENTS.md, an OpenAPI spec, MCP manifests and registry presence, crawlable sitemaps — that let an agent find a usable version of the documentation on a plain HTTP fetch.
Comprehensioncomprehension: Comprehension is whether an AI agent can operate an API from its documentation: find the facts an integration depends on, construct correct requests, and refuse to invent capabilities the API does not support.
Behavioral measurementbehavioral-measurement: Behavioral measurement grades documentation by what models actually do with it: models are given the docs as their only source, quizzed on tasks a real integration depends on, and their answers are graded mechanically against ground truth cited to the documentation itself.
Task banktask-bank: A task bank is a versioned, published set of tasks used to quiz models on an API's documentation.
Trap tasktrap-task: A trap task asks a model to do something the API does not support; the correct answer is a refusal.
Closed-book baselineclosed-book-baseline: A closed-book baseline runs the same measurement tasks with no documentation provided, establishing what a model already knows about an API from training.
Docs liftdocs-lift: Docs lift is the improvement a model shows when given an API's documentation, compared with the closed-book baseline — the same tasks run without the docs.
Capability floorcapability-floor: The capability floor is the lowest model tier that can reliably operate an API from its documentation.
Tokens-to-comprehensiontokens-to-comprehension: Tokens-to-comprehension is the amount of documentation a model must read before it can correctly operate an API — the reading cost of understanding, denominated in the tokens an agent spends ingesting docs.

Adjacent disciplines

AEO (Answer Engine Optimization)aeo: Answer Engine Optimization (AEO) is the practice of structuring content so AI answer engines — chat assistants and AI-powered search — can find, understand, and cite it when composing answers.
GEO (Generative Engine Optimization)geo: Generative Engine Optimization (GEO) is the practice of making content visible and citable in the outputs of generative engines — systems that synthesize answers with AI rather than returning ranked links.
ADO (Agent Discovery Optimization)ado: Agent Discovery Optimization (ADO) is the emerging discipline of making an API discoverable and legible to AI agents: publishing machine-readable entry points such as llms.txt, AGENTS.md, an OpenAPI spec, and MCP surfaces, and serving documentation that renders on a plain HTTP fetch.
SEO for agentsseo-for-agents: SEO for agents is the informal umbrella term for optimizing a site or API so AI agents can find and use it.
Agent trafficagent-traffic: Agent traffic is the portion of requests to a website or API that originates from AI systems rather than humans in browsers — crawlers gathering training or retrieval data, assistants fetching pages to answer questions, and agents reading documentation while executing tasks.