A
agency
Definition A way of working with AI where you set it up to act on your behalf and let it run. You shape what it knows and how it behaves, rather than directing each step.
Use Agency suits work you've done often enough to describe well, like a weekly report or sorting an inbox. Give it clear limits and a way to check its work, and review what it did before you trust the next run.
See also agent, automation, augmentation, permission
agent
Definition An AI that takes actions, not just answers: it opens files, uses tools, searches, and decides its next step until the task is done.
Use An agent works in a loop: read the situation, pick an action, look at the result, repeat. Its work is only as good as what it's given at the start and what it's allowed to touch. Watch the first few runs closely.
See also subagent, tool use, permission, hook
agent harness
Definition The program a model runs inside. It gives the model its loop, its tools, and its permissions, so the same model can act differently in a different harness.
Use When an assistant does something you didn't expect, ask whether it was the model or the harness: what it could see, which tools it had, what it was allowed to do. Each coding agent, like Claude Code, Codex, or Gemini CLI, is a harness around its company's models.
Also called harness
See also agent, model, tool use, permission, coding agent
AGI
Definition Short for artificial general intelligence: an AI that could learn and reason across almost any task as well as a person, or better. The term is debated, since people mean different things by "human-level," and there is no accepted test for it.
Use When a company or a headline says a system has reached AGI, or is close, ask what test they mean; with no agreed test, the claim is hard to check. Today's systems are uneven: the same ones that win math contests still misread an analog clock.
Also called artificial general intelligence, human-level AI
See also artificial intelligence, jagged intelligence, benchmark
AI assistant
Definition A program you talk to in plain language that writes, explains, plans, and answers questions. ChatGPT, Claude, and Gemini are examples.
Use How useful an assistant is depends less on which one you pick and more on what it knows about you and your work at the start of each session.
See also chatbot, model, session, persistent memory
AI disclosure
Definition Telling the people who receive your work what the AI did, and what you checked yourself.
Use A short line is often enough: "Drafted with an AI assistant; the figures checked against the source." Expectations differ by setting, and school, work, and clients each have their own.
Also called transparency (the 4Ds)
See also output evaluation, 4Ds
AI solution architect
Definition A person who designs how AI fits into someone else's work, end to end: what it does, what it connects to, who checks it, and how success is measured.
Use The title is still settling, and what it covers varies by employer. Expect the work to start from the problem, not the tool. Example: Mapping a clinic's intake process before choosing which steps an agent handles.
Also called AI solutions architect
See also brief, human in the loop, ownership, verification contract
API
Definition A way for one program to talk to another. An AI company's API lets your own software send requests to its models.
Use An API is usually paid by use, separately from a chat subscription, and needs an API key kept safe. Most people never need one; people building their own tools do.
Also called application programming interface
See also API key, tool use
API key
Definition A long secret code that proves a request comes from your account. Anyone who has it can spend your money.
Use Keep keys in a password manager, never in a document, a chat, or a file you share. If one leaks, revoke it and make a new one.
See also API, permission
artificial intelligence
Definition Computers doing tasks that usually take human intelligence, like understanding language, recognizing images, and making decisions. Most AI today works by finding patterns in large amounts of data.
Use "AI" covers a spam filter and a chat assistant alike, so when a headline or a sales pitch says "AI," ask which kind. Some AI is narrow, good at one task; the large language models behind chat assistants are general-purpose and handle many tasks, unevenly. The term dates from 1955.
Also called AI
See also machine learning, generative AI, large language model, AGI
audit log
Definition A record of who did what, and when.
Use When several assistants or people work on one job, the log is how you find where something went wrong. Keep it somewhere a stranger could read and follow it.
See also ownership, handoff
augmentation
Definition A way of working with AI as a partner: you think and build together, each doing the part you're better at.
Use Most good everyday use is augmentation. You bring the judgment and the context; the assistant drafts, checks, and suggests.
See also automation, agency, iteration
automation
Definition Having AI do a specific task the way you told it to, usually the same way every time.
Use Automate a task only after you've done it by hand and know what "right" looks like. Otherwise you automate the mistakes too.
See also augmentation, agency, runbook
B
batching
Definition Grouping similar tasks and doing them together in one pass, so you pay the cost of switching between kinds of work once instead of every time.
Use Hand an assistant a batch, not a trickle: twenty files renamed under one rule beats twenty separate asks, and you check the rule once. A workflow is the same idea at a larger scale: a chain of steps you set up once and run as one batch of work. With an API the word means something close: sending many requests at once, answered later and usually for less.
Also called task batching; batch processing (API)
See also digital chores, brief, API, automation
benchmark
Definition A standard test for AI: a fixed set of problems and a way to score the answers, so different systems can be measured and compared on the same task.
Use Benchmarks show real progress, and they can mislead: a model can be tuned to the test, and the test may not look like your work. A system can top one benchmark and stumble on an easy task somewhere else. Try a model on your own task before you trust a score.
See also jagged intelligence, model, output evaluation
brief
Definition A short written description of a task: what it's for, what done looks like, and what's out of scope.
Use A brief written before you start saves more time than any prompt trick. Hand the same brief to a person or an assistant; if either is confused, the brief needs work.
See also definition of done, prompt, system prompt
C
chatbot
Definition A program you hold a conversation with, by typing or by voice. Older chatbots matched keywords, like a phone menu; today's AI chatbots write their own replies to almost any question.
Use Surveys and news say "chatbot," the companies say "assistant," and they mostly mean the same tools. An AI chatbot writes each reply by predicting likely words, which is why it reads fluently and can still be wrong, and why the same question can get a different answer.
Also called AI chatbot, chat assistant
See also AI assistant, large language model, hallucination, non-determinism
checkpoint
Definition A saved point in your work that you can go back to if something breaks.
Use Save one before any big change. With version history, checkpoints cost nothing, so make many.
See also version history, regression
coding agent
Definition An AI tool that works with the files on your computer: it reads them, changes them, and runs commands, with your permission. It isn't only for code.
Use You open it on a folder; it works inside that folder and asks before it reaches outside. Depending on the tool, it runs in a terminal, in a code editor, or in the company's desktop app.
Also called agentic coding tool; Claude Code, Codex, Gemini CLI (examples)
See also agent, agent harness, permission, plan mode, hook
compaction
Definition Shrinking a long conversation into a summary so the assistant has room to keep working.
Use A summary keeps the gist and can drop the details: numbers, names, decisions. Before compacting, write the facts that matter to a file. Coding agents do it on request, and on their own when the window fills.
See also context window, handoff, persistent memory
connector
Definition A link that lets an AI assistant read from, or act in, another app, like your email, calendar, or files.
Use Connect only what the task needs, and check what each connector can do; read-only is safer than read-and-write. Anything the assistant reads through a connector, an email or a web page, can also try to give it instructions.
See also MCP, permission, tool use
context window
Definition Everything the AI can see at one time: your messages, its replies, and any files it has read. When the session ends, it's gone.
Use The window has a limit. Long sessions fill it, and the assistant starts losing early details or getting vaguer. Keep what matters in files, and start a fresh session from a summary when one drifts.
See also persistent memory, compaction, token, session
counterfactual
Definition What would have happened without something: the comparison you need before you can say what a tool, a person, or a change actually added.
Use Ask it of your own AI work: what shipped that you could not have built alone, in the time you had? That difference is leverage. A job you'd have done just as well by hand, only slower, is a smaller win than one you'd never have finished.
See also augmentation, benchmark, output evaluation