Glossary
AI agent glossary
95 essential terms for working with and building AI agents, explained in plain English and linked to the lessons that go deeper.
A
Agent harness
The software wrapped around a language model that runs the agent loop, supplies tools, manages context, and enforces permissions.
Agent hooks
Your own scripts that an agent harness runs automatically at set points, such as before a tool call or when the agent finishes, to check or act.
Agent loop
The repeating cycle at the heart of every AI agent: the model picks an action, a tool runs it, and the result is fed back until the task is done.
Agent memory
Any mechanism that lets an agent keep useful information beyond a single conversation, such as notes files, databases, or retrieved past sessions.
Agent SDK
A software library that packages the agent loop, tool handling, and common features so you can build agents without writing that plumbing yourself.
Agent skills
Reusable folders of instructions, scripts, and reference files that an agent loads only when a task calls for them.
Agent2Agent (A2A) protocol
An open protocol that lets AI agents built by different vendors discover each other, exchange messages, and delegate tasks across systems.
Agentic AI
AI systems that pursue goals by planning and taking actions with tools over multiple steps, rather than just answering a single prompt.
Agentic coding
Writing software by delegating multi-step tasks to a coding agent that reads files, edits code, and runs commands, while you direct and review.
AGENTS.md and CLAUDE.md
Plain markdown files in a project that coding agents read automatically to learn its conventions, commands, and rules.
AI agent
A system where a language model works toward a goal by deciding which actions to take, using tools, and checking results in a loop.
Autonomous agent
An agent that carries out a task from start to finish with little or no human input along the way.
B
Background agent
An agent that runs on its own, often in the cloud or on a schedule, while you do other work, then reports back when it finishes.
Batch API
A way to submit many model requests at once for asynchronous processing, trading immediate responses for lower prices and higher throughput.
Browser agent
An agent that operates a web browser on your behalf: opening pages, reading them, clicking, typing, and filling in forms.
C
Chain of thought
Having a model reason step by step in writing before giving its answer, which tends to improve results on multi-step problems.
Chunking
Splitting long documents into smaller passages before embedding and indexing them, so retrieval can return just the relevant parts.
Claude Agent SDK
Anthropic's library for building agents on the same foundation that powers Claude Code, available for TypeScript and Python.
Claude Code
Anthropic's agentic coding tool that reads your codebase, edits files, runs commands, and completes programming tasks from natural-language requests.
Coding agent
An AI agent specialized for software work: it reads code, edits files, runs commands and tests, and iterates until a programming task is done.
Compaction
Summarizing the older parts of a long agent conversation so the work can continue without overflowing the context window.
Computer use
The ability of a model to operate a computer like a person does, by looking at screenshots and sending mouse clicks and keystrokes.
Context engineering
The practice of deciding exactly what information goes into a model's context window at each step so an agent has what it needs and nothing that distracts it.
Context rot
The tendency of a model's accuracy and instruction-following to degrade as its context fills with more tokens, even well below the window limit.
Context window
The maximum amount of text, measured in tokens, that a model can consider at once, including instructions, conversation history, and tool results.
Cost per task
The total spend, mainly model tokens plus tool and infrastructure costs, needed for an agent to complete one unit of useful work.
D
Data exfiltration
Sensitive data being sent out of a system without authorization; for agents, usually by tricking the model into leaking it through a tool or link.
Deep research
An agent mode that researches a question for many minutes, running many searches and reading many sources, then writes a cited report.
Dev container
A development environment defined in a config file and run inside a container, often used to give coding agents an isolated, reproducible workspace.
Durable execution
Running a long, multi-step process so its progress is saved at each step and it can resume after a crash, restart, or timeout instead of starting over.
E
Embeddings
Lists of numbers that represent the meaning of text, so that pieces of text with similar meanings end up close together and can be found by similarity.
Evals
Repeatable tests that measure how well an AI system performs on a set of representative tasks, so you can tell whether a change made it better or worse.
F
G
Git worktree
A Git feature that gives one repository several working folders, each on its own branch, so multiple agents can work in parallel without colliding.
Golden dataset
A curated set of realistic inputs paired with verified correct or reference outputs, used as the trusted benchmark for evaluating an AI system.
Grounding
Tying a model's answers to specific, trusted sources such as documents, search results, or tool outputs, rather than relying on what it remembers.
Guardrails
Checks and limits placed around an AI system to keep its inputs, actions, and outputs within safe and acceptable bounds.
H
Hallucination
When a model states something false or made up with the same confidence as a fact, such as an invented citation, function, or statistic.
Handoff
In a multi-agent system, passing control of a conversation or task from one agent to another that is better suited to continue it.
Headless mode
Running an agent non-interactively from a script, CI job, or scheduler, with the task passed in and the result printed or saved instead of a live chat.
Human-in-the-loop
Designing an agent so a person reviews, approves, or corrects its work at chosen points instead of letting it act entirely on its own.
I
J
K
L
Large language model (LLM)
An AI model trained on vast amounts of text to predict what comes next, which lets it write, summarize, reason, and follow instructions.
Latency
How long a user waits for a response, which for agents adds up across every model call and tool call in the loop.
Least privilege
The security principle of giving an agent only the permissions and access it truly needs for its task, and nothing more.
Lethal trifecta
The dangerous combination of an agent having private data, exposure to untrusted content, and a way to send data out, which makes data theft possible.
LLM-as-judge
Using a language model to grade the outputs of another model or agent against a rubric, so quality can be measured at scale.
M
Max tokens
A request parameter that caps how many tokens the model may generate in one response; it limits output length, not the size of the input.
MCP client
The part of an AI application that connects to MCP servers, discovers their tools and resources, and passes them to the model.
MCP Inspector
An official developer tool for testing and debugging MCP servers in a browser, letting you list and call tools, resources, and prompts by hand.
MCP resources
Read-only data an MCP server exposes, such as files, records, or documents, each identified by a URI and supplied to the model as context.
MCP server
A program that exposes tools, data, or prompts to AI applications through the Model Context Protocol, so any compatible agent can use them.
MCP transport (stdio and Streamable HTTP)
How an MCP client and server exchange messages: stdio for local servers launched as a subprocess, or Streamable HTTP for remote servers.
Message roles (system, user, assistant)
The labels that mark who said each part of a conversation sent to a model: system instructions, the user, or the assistant itself.
Model Context Protocol (MCP)
An open standard for connecting AI applications to external tools and data, so one integration works across many different agents.
Model routing
Sending each request or step to the model best suited to it, such as a fast, cheap model for simple work and a more capable one for hard problems.
Multi-agent system
An application where several AI agents, each with its own instructions, tools, or context, work together to complete a larger task.
O
Open-weight model
A language model whose trained weights are published for anyone to download and run on their own hardware, unlike models available only through an API.
Orchestrator-worker pattern
A multi-agent design where a lead agent breaks a task into pieces, hands each to a worker agent, and combines their results.
P
Parallel tool calls
When a model requests several tool calls in a single response, so your code can run them at the same time and return all the results together.
pass@k
An eval metric giving the chance that at least one of k attempts at a task succeeds, used to measure models and agents whose outputs vary.
Permission modes
Settings that control how much an agent may do without asking, from approving every action to editing files freely to running fully unattended.
Plan mode
A coding agent setting where the agent may read and explore but not edit files, so it proposes a plan for you to approve before making changes.
Prompt caching
Reusing the already-processed beginning of a prompt across requests, which makes repeated calls with the same prefix faster and cheaper.
Prompt chaining
Splitting a task into a fixed sequence of model calls, where each call's output becomes input to the next, often with checks in between.
Prompt engineering
The craft of writing instructions and examples that reliably get a language model to produce the output you want.
Prompt injection
An attack where instructions hidden in content an agent reads, like a web page, email, or file, trick it into doing something its user never asked for.
R
Rate limits
Caps a model provider places on how many requests or tokens you can use per minute or day; exceeding them returns errors until the window resets.
ReAct
A pattern where a model alternates between reasoning about what to do and acting with a tool, using each observation to guide the next step.
Reasoning model
A language model trained to think through problems step by step before answering, trading extra time and tokens for better results on hard tasks.
Red teaming
Deliberately attacking your own AI system, the way an adversary would, to find unsafe behavior and security holes before real users do.
Reranking
A second retrieval stage that rescores an initial set of search results with a more accurate model, so the best matches end up at the top.
Retrieval-augmented generation (RAG)
Finding relevant documents for a question and adding them to the model's prompt, so it can answer from your data instead of memory alone.
S
Sandboxing
Running an agent inside an isolated environment, such as a container or virtual machine, so any mistakes or attacks stay contained.
Semantic search
Search that finds results by meaning rather than exact words, usually by comparing embeddings of the query and the documents.
Slash commands
Shortcuts typed with a leading slash in an agent's chat, either built in, such as clearing context, or custom ones that expand into saved prompts.
Small language model (SLM)
A language model with far fewer parameters than frontier models, making it cheaper and faster, and often small enough to run on a laptop or phone.
Stop reason
A field in a model's response that says why it stopped generating, such as finishing its turn, requesting a tool, or hitting the token limit.
Streaming
Receiving a model's response piece by piece as it is generated, so users see text appear immediately instead of waiting for the whole reply.
Structured output
Getting a model to return data in a precise, machine-readable format such as JSON that matches a schema, rather than free-form text.
Subagent
A helper agent launched by a main agent to handle a focused subtask in its own fresh context, returning only a summary of the result.
System prompt
The standing instructions given to a model before the conversation starts, setting its role, rules, tone, and how it should use its tools.
T
Temperature
A setting that controls how random a model's word choices are: lower values give more predictable output, higher values more varied output.
Token
The unit of text a language model reads and writes, usually a word or part of a word, and the basis for context limits and pricing.
Tool poisoning
An attack where malicious instructions are hidden in a tool's name, description, or schema so the model follows them when it reads the tool list.
Tool result
The message your code sends back to the model after running a tool it requested, carrying the output or an error so the agent can continue.
Tool use (function calling)
The ability of a language model to request that your code run a specific function, like a search or database query, and then use the result.
Tracing and observability
Recording every model call, tool call, input, and output an agent makes, so you can see exactly what happened and why when things go wrong.
V
Vector database
A database designed to store embeddings and quickly find the ones most similar to a query, used to power semantic search and RAG.
Vibe coding
Building software by describing what you want to an AI and accepting its code largely without reading it, judging only whether the result seems to work.