Agent
Ukrainian: агент
A model in a harness, calling tools in a loop to reach a goal. A fixed pipeline whose steps are coded in advance is a workflow, not an agent.
99rabbits / Glossary
Glossary
These are the words I use when I write about AI agents, the tools around them, and how offers and payments reach them: one term per concept, with the words it replaces. I checked every term against named sources; both show under each entry when you switch on Details. The Ukrainian beside each term is for my Ukrainian readers; where I found no Ukrainian source, the rendering is marked as a proposal.
Ukrainian: In Ukrainian, a loanword gets the English in brackets on first use, as in «харнес (harness)»; names such as MCP, x402 and CLAUDE.md stay in Latin script.
Agents and the code around them
Harness
Ukrainian: харнес
Everything around the model that turns it into an agent: tools, system prompt, what gets loaded into context, permissions, hooks.
Agent loop
Ukrainian: агентний цикл
The cycle the harness runs: the model proposes a tool call, the harness executes it, the result goes back to the model, repeat.
Turn
Ukrainian: хід(turn)proposal
One user message and everything the agent does before it hands back.
Session
Ukrainian: сесія
One run of work with an agent, from an empty context window until it is cleared, closed or compacted. A new session starts without the history of the last one.
Context window
Ukrainian: контекстне вікно
Everything the model sees on one request: the system prompt, the conversation so far and every tool result fed back in. It is finite, and what is outside it does not exist for the model.
Token
Ukrainian: токен
The unit a model reads and writes, roughly a word or a piece of one. The size of a context window, the price and the latency are all counted in tokens.
Tool
Ukrainian: інструмент
A function the harness exposes for the model to call, such as reading a file, running a command or searching the web. The agent sees and changes the world outside the model only through tools.
Tool call
Ukrainian: виклик інструмента
The model’s request to run a tool: the tool’s name and its arguments, written as structured output. The harness decides whether to run it, then sends back the tool result.
Tool result
Ukrainian: результат виклику інструмента
The data a tool call returns to the model.
MCP
Ukrainian: MCP
The Model Context Protocol, an open standard for connecting a host to tool servers. The agent never calls MCP itself; it calls a tool that the host got from an MCP server.
Prompt
Ukrainian: промпт
The message a user sends the model; it opens a turn.
Response
Ukrainian: відповідь
What the model writes back to the user at the end of a turn. Anthropic’s API and Pocock also call a tool call a response, so write “final response” where the difference matters.
System prompt
Ukrainian: системний промпт
The instructions the harness sends ahead of the conversation on every request. The vendor or the operator writes it, not the user, and models are trained to rank it above the user’s messages.
Instruction
Ukrainian: інструкція
Text that tells the model what to do, in a system prompt, a CLAUDE.md or a skill. A safeguard written as an instruction holds only while the model complies.
Subagent
Ukrainian: субагент
An agent another agent spawns for a sub-task; it runs in its own context window and reports one result back.
Orchestrator
Ukrainian: оркестратор
The agent that splits a task, hands the pieces to subagents and synthesises what comes back.
Skill
Ukrainian: скіл
A folder with a SKILL.md that the agent loads when a task matches it.
Hook
Ukrainian: хук
A command the harness runs on its own at a set point, such as after every tool call, without the model deciding to.
Effort
Ukrainian: рівень міркування(effort)
How much a model reasons before it answers: “effort level” in Claude Code, “reasoning effort” in Codex.
Who runs the agent and what the user sees
Host
Ukrainian: хост
The application that runs the model connection and shows the response: a chat app, an IDE, or a bot an operator runs. In a coding agent the host and the harness are one program: “host” names its MCP side and what it shows the user.
Platform
Ukrainian: платформа
A chat product run by a model company, such as ChatGPT or claude.ai, that hosts other people’s tools and can sell its own ads.
Operator
Ukrainian: оператор
Whoever runs an agent product for users and controls what its host adds to a response.
User
Ukrainian: користувач
The human who reads the response.
MCP App
Ukrainian: MCP App
Interactive UI a tool returns that the host renders inside the conversation.
Claude Code
Permission prompt
Ukrainian: запит дозволу
What the harness shows the user before it runs a tool call that no rule allows.
Allow rule
Ukrainian: правило дозволу(allow)
An entry under permissions.allow in Claude Code’s settings that lets matching tool calls run without a permission prompt.
Global CLAUDE.md
Ukrainian: глобальний CLAUDE.mdproposal
~/.claude/CLAUDE.md, the file that loads into every Claude Code session. Anthropic’s docs call it “User instructions”; Claude Code’s own label for it says “global”.
Advisor
Ukrainian: радникproposal
Claude Code’s advisor tool: a second, usually stronger model that the main model consults when it decides to.
Where an offer can enter
Entry pointmy term
Ukrainian: точка входуproposal
A place in an agent’s flow where someone can attach a sponsored offer. There are seven: feed item, webhook, tool result, MCP App, host line, platform ad, and a priced offer an agent can pay for.
Ingestion
Ukrainian: надходженняproposal
How content reaches the agent before a run: webhooks, RSS, email, polling an API.
Webhook proxy
Ukrainian: вебхук-проксіproposal
A middleman that receives webhooks and forwards them, and so can add to them.
The model’s gatemy term
Ukrainian: фільтр моделіmy term
The boundary in an agent’s flow above which an offer reaches the user only if the model repeats it. Offers attached below it, by the host or the platform, reach the user every time.
Surface
Ukrainian: поверхняproposal
Whatever finally shows an offer: a user’s screen, or an agent’s own decision.
Untrusted content
Ukrainian: ненадійний контент
Text from third parties that reaches the model’s context, such as a feed item, a web page or a tool result. A sponsored item arriving this way is untrusted content, and Anthropic’s docs file instructions inside it under indirect prompt injection.
Prompt injection
Ukrainian: промпт-ін’єкція
An attack that hides instructions in untrusted content, so that the model follows them as if the user or the operator had given them. Willison coined the term in 2022 for mixing trusted and untrusted text in one context.
Money
Settlement
Ukrainian: розрахунок
The event that moves money: a click followed by a sale reported back, or a payment the agent makes itself.
x402
Ukrainian: x402
An open standard for paying over HTTP: the server answers 402 Payment Required with its terms, and the client retries with a signed payment.
MPP
Ukrainian: MPP
The Machine Payments Protocol, an open standard for agents to pay, co-authored by Tempo and Stripe, which works with stablecoins, cards and bank transfers.
Facilitator
Ukrainian: фасилітатор
In x402, the service that verifies a payment and settles it onchain for the seller, without holding the funds.
Postback
Ukrainian: постбек(affiliate jargon)
A server-to-server report of a conversion, such as a sale, from the merchant to the affiliate network. Plain-English form: “a sale reported back”.
Browsers
Challenge page
Ukrainian: сторінка перевіркиproposal
The page a bot-protection service puts in front of the real one, such as Cloudflare’s “Just a moment...”.
Clearance cookie
Ukrainian: кукі cf_clearance
Cloudflare’s cf_clearance cookie, proof that a browser passed a challenge.
Data directory
Ukrainian: каталог даних(user data directory)proposal
The folder that holds all of a Chrome install’s profiles. Chrome 136 and later refuse a debug port on the default one.
Ref
Ukrainian: ref
agent-browser’s handle for an element in a snapshot of the page’s accessibility tree, such as @e3.
Coinages
Interceptormy term
Ukrainian: перехоплювачproposal
An agent that rewrites its human’s message before it is sent and stays invisible in the conversation.
Demonmy term
Ukrainian: демонmy term
An agent that speaks in the conversation on its own line, next to its human; a personal spirit in the classical sense.
Ways of working
Grilling
Ukrainian: допит(grilling)my term
A way to develop a plan or a draft with an agent: the agent interviews the user one decision at a time and proposes an answer to each, until nothing is left undecided.
Learning
Bloom’s taxonomy
Ukrainian: таксономія Блума
A framework for classifying what students are meant to learn, from a 1956 handbook edited by Benjamin Bloom. Its 2001 revision runs from Remember to Create; the 1956 original ran from Knowledge to Evaluation.