Learn the words used to explain AI and build software with an agent. Start with a definition, then follow the idea into a lesson or a practical decision.
164 of 164 terms
Also called steering vector, persona vector
Changing a model's behavior by nudging its internal activations directly at runtime — no retraining, no prompt changes.
Also called agent credentials
Managing agent identity and credentials under least privilege.
Also called CI agents
Running versioned agent workflows from delivery infrastructure.
Also called code interpreter
Running model-generated commands or programs under controlled authority.
Also called constraint-based prompting
Boundaries that define what an agent may do and what success must satisfy.
Also called AI data governance, agent retention
Policies and controls for data collection, use, storage, location, retention, and deletion.
Also called agent UX, AI UX
Interface design that makes agent work understandable, steerable, and recoverable.
Also called agent debugging
Tracing a bad outcome to the system component that produced it.
Also called harness
The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Also called hook
A script the harness runs automatically at a fixed point in an agent's loop — before a tool call, after an edit, on session start.
Also called AI incident response
The operating process for containing and learning from harmful agent behavior.
Also called agent economics
Managing time and spend across model calls, tools, retries, and concurrency.
Also called reason-act-observe loop
A cycle in which an agent observes, decides, acts, and interprets the result.
Also called long-term agent memory
Information an agent can persist and retrieve beyond the immediate context.
Also called coding agent
A mode in which one agent can inspect, edit, run commands, and iterate toward an outcome.
Also called agent tracing
The ability to inspect an agent's actions, state, outputs, and resource use.
Also called multi-agent orchestrator
A coordination layer that assigns work and manages state across agent workers.
Also called plan mode
Turning a goal into an inspectable sequence that may change as evidence appears.
Also called agent monitoring
Tracking agent quality and operations after deployment.
Also called concurrency limits
Controls that bound request volume and simultaneous work.
Also called behavioral regression suite
Repeating stable behavior checks after model, prompt, tool, or retrieval changes.
Also called sandboxed environment
An isolated runtime with explicit limits for agent-executed work.
Also called agent threat model
Systematically identifying how an agent could be misled, abused, or given excessive authority.
Also called session hygiene
Keeping agent work coherent across growing history, compaction, handoffs, and restarts.
Also called skill, SKILL.md
A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
Also called work decomposition
Splitting work into bounded units that can be owned and verified independently.
Also called multitenant agent isolation
Enforced separation of customer data and authority across every agent subsystem.
Also called tool calling
Giving a model structured actions for reading or changing external systems.
Also called agent security boundary
A separation that limits what an agent or untrusted input can access or change.
Also called tool loop, agent loop
The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.
Also called software factory
A repeatable system connecting specifications, agents, verification, review, and delivery.
Also called agent
An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
Also called self-attention, attention mechanism
The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
Also called code completion
A model predicts the next code while the developer remains the author.
Also called executable verification
Using executable checks to judge whether agent-produced work meets its contract.
Also called agent swarm
Multiple agents that coordinate work with limited direct human or central control.
Also called asynchronous agent
An agent that continues scoped work without a person attending the live loop.
Also called agent feedback constraint
Signals and constraints that force an agent to stop or correct a failing path.
A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Also called browser agent
Programmatic operation and inspection of a rendered web interface.
Also called calibrated
How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
Also called CoT
Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.
Also called AI chat
A conversational model that answers questions without directly owning repository changes.
Also called text splitting
Splitting documents into passages small enough to embed and retrieve individually — the step that quietly determines whether retrieval works at all.
Also called terminal agent
An agent operated from a command line with access to repository tools and shell commands.
Also called command-line tool
A command-line capability an agent can discover and compose through shell commands.
Also called model catalog
A selection of first-party and partner models offered inside a larger cloud platform's security and billing boundary.
Also called compounding workflow
Improving the environment after each task so future agent work becomes more reliable.
Also called compaction, context compression
Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
Deciding what an agent sees at each step — the discipline that replaced prompt engineering once context windows got large enough to fill badly.
Also called context length, context
The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Also called regional data residency
A requirement that data be processed or stored within a specified country or geographic region.
Also called relational database, SQL
Persistent structured storage queried and changed through explicit operations.
Also called dedicated deployment
A model deployment with capacity reserved for one customer instead of shared across many users.
Also called agent delegation, agent authorization
A traceable, limited identity an agent uses on behalf of a sponsor.
Also called reproducibility
Whether the same input reliably produces the same output — something LLM systems mostly lack, which changes how you test and debug them.
Also called first-party model API
An API operated by the same lab that creates the model family it serves.
Also called distilled
Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
Also called chunking
Splitting source material into units that can be indexed and retrieved.
Also called durable execution
Persisting agent state so long-running work can resume after interruption.
Also called embeddings, vector embedding
A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
Also called evals, evaluation
A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
Also called eval set, golden set
A repeatable set of representative cases used to measure system behavior.
Also called fallback routing
Automatically sending a failed request to another model endpoint or provider so the application can keep working.
Also called finetuning, fine-tune
Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
Also called top models, state of the art models
The changing set of models that lead meaningful workloads at a point in time.
Also called GEO, AEO, answer engine optimization
Making a site the source an AI assistant cites when answering, rather than a link in a list of ten blue results.
Also called Git, version control
A system for recording and comparing changes to a codebase.
Also called worktree
A separate working directory attached to the same Git repository.
Also called grounding
Generation tied to supplied evidence rather than unsupported model memory.
Also called grounded
Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
Also called hallucinate
When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
Also called agent harness
Designing the full environment and feedback system around an agent.
Also called autonomous agent run, background agent
An agent running with no one watching — in CI, on a schedule, or triggered by an event — with its output reviewed after the fact.
Also called web API, REST API
A common way for software to request data or actions from another service.
Also called human in the loop, human on the loop
Defined points where a person approves, monitors, or retains a judgment.
Also called approval flow, agent interruption
A structured transfer of control and state between an agent and a person.
Also called HITL
Requiring a person's approval at specific points in an automated process, chosen so the irreversible steps are the ones a human sees.
Combining keyword and semantic search in one ranked result, so exact terms and paraphrases both find what they should.
Also called independent agent review
Using a separate role or model instance to evaluate an implementer's work.
Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Also called AI inference cloud
A cloud platform specialized in running open or custom AI models efficiently at production scale.
Also called model provider, LLM API provider
A service or runtime that executes a model and exposes it to an application.
Also called editor edit
A model-authored change applied directly to a selected region or file.
A prompt crafted to make a model ignore its own guidelines — usually through roleplay, hypotheticals, or encoding rather than a direct request.
Also called training cutoff
The date after which a model saw no training data — everything later has to reach it through search, tools, or the context window.
Also called KV caching, key-value cache
The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
Also called LLM
A neural network trained to predict and generate sequences of tokens.
Also called large language model, language model
A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
Also called agent evaluation, evals
Measuring model or agent behavior against an explicit success definition.
Also called model-graded eval, AI judge
Using one model to score another's output against a rubric, so quality can be measured at a scale human grading cannot reach.
Also called llms txt
A file at a site's root that gives AI systems a curated, plain-text map of its most useful content — robots.txt's counterpart for comprehension rather than permission.
Also called low-rank adaptation
A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
Also called frontier lab, AI lab
An organization that trains and releases important foundation-model families.
Also called Model Context Protocol
The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
Also called MoE, mixture of experts
A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
Also called fine-tuning decision, prompting versus RAG
Changing the system or model so it performs a target workload more reliably.
Also called MCP
An open protocol for connecting AI applications to tools and data sources.
Also called model lineage
A related line of models released under a shared name and technical lineage.
Also called inference
Running a trained model to produce an output.
Also called AI router, LLM router
A service that sends model requests across multiple labs or hosts through one API.
Also called model cascade, model fallback
Sending each request to a model chosen by the difficulty of the task, rather than using one model for everything.
Also called model choice
Choosing a model using workload evidence and operating constraints.
Also called multi-agent system, subagent
Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
Also called multimodality
A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
Also called vision pipeline, media AI pipeline
A typed flow that turns media into model-ready evidence and verifiable outputs.
Also called multimodal model
The ability to understand or generate more than one kind of media.
Also called open-weight, open source model
A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Also called open model, local model
A model whose trained weights can be downloaded and run outside the maker's hosted API.
Also called OpenAI compatibility, OpenAI-compatible
An API that accepts the common OpenAI request shape so existing clients can call another provider with fewer code changes.
Also called multi-agent parallelism
Multiple isolated agents working simultaneously on independent work streams.
Also called planning mode
A read-only agent mode that investigates and proposes a plan without editing anything, so you approve the approach before any code changes.
Also called foundation model
A model trained broadly before it is adapted or prompted for a particular application.
Also called pre-training
The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
Also called AI use-case selection, product judgment
Choosing a real user problem where an AI system creates verifiable value at acceptable cost and risk.
Also called coding basics
The core ideas used to express behavior in source code.
Also called prompt structure
The instructions, context, examples, and output contract inside a model request.
Also called context caching
Reusing the model's processed form of a repeated prompt prefix so subsequent calls skip re-reading it, cutting cost and latency substantially.
An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.
Also called quantized
Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
Also called retrieval-augmented generation
Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.
Also called voice agent, realtime agent
An agent that processes streaming audio or visual input while a session is live.
Also called reasoning budget, test-time compute
The inference-time budget a reasoning model spends before answering.
Also called thinking model
A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
Also called adversarial testing
Deliberately attacking your own AI system to find what makes it fail before someone else does.
Also called AGENTS.md, project instructions
Versioned guidance that tells agents how a repository is structured and verified.
Also called cross-encoder reranking
A second pass that re-scores retrieved candidates by reading each one against the query, fixing the ordering that fast vector search got approximately right.
Also called safe retries
Recovering from failure without duplicating a completed side effect.
Also called information retrieval
Selecting evidence relevant to the current request.
Also called answer-time fetch, AI search crawler
A bot that fetches a page at the moment someone asks a question, so an assistant can cite it — distinct from a crawler collecting training data.
Also called reranker
Reordering retrieved candidates with a stronger relevance model.
Also called RAG
Retrieving external evidence and placing it in context before generation.
Also called reinforcement learning from human feedback
Reinforcement learning from human feedback — training a model to prefer answers humans rate as better, which turns a raw text predictor into a usable assistant.
Also called temperature, top-p
Parameters that shape how deterministic or varied model generation becomes.
Also called sandboxing, sandboxed
An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
Also called self-optimizing agent
A system that proposes and evaluates changes to its own operating components.
Also called serverless model API
Model inference that scales on demand without the developer reserving or managing a continuously running server.
Also called multi-agent memory
Governed durable knowledge shared among multiple agent workers.
A named, reusable prompt invoked with a leading slash, turning a workflow you repeat into one word you type.
Also called SDD, spec-first
Writing and agreeing the specification first, then having an agent implement against it — the spec, not the code, becomes the artifact you maintain.
Also called speculative sampling
A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
Also called token streaming
Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
Also called JSON mode, constrained decoding
Forcing a model's response to match a schema, so downstream code can parse it instead of guessing at prose.
Also called sub-agent
A separate agent spawned by another to do one scoped piece of work in its own context, returning only the result.
Also called SWE-bench Verified
The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
Also called agent task brief
Expressing an outcome, context, constraints, and definition of done for an agent.
Also called command line, shell
A text interface for navigating files and running programs.
Also called inference-time compute, inference-time scaling
Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.
Also called software delivery
The checks and release steps that turn source changes into running software.
Also called tokens per second
How much model output a provider can generate over time, commonly measured in output tokens per second.
Also called TTFT
The delay between sending a model request and receiving the first piece of its response.
Also called tokens
The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
A cap on how many tokens a task, session, or agent run may consume — the practical control on both cost and how long an agent will grind.
Also called permission model, allowlist
The rules governing which tools an agent may call and which need confirmation — the boundary between a mistake and an incident.
Also called function schema design
Designing a tool interface so a model can select and call it correctly.
Also called function calling, tool calling
A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
Also called semantic search
Finding content by similarity between embeddings.
Also called vibecoding
Building software by describing what you want to an AI in plain language and steering the result, instead of writing every line yourself.
Using coding agents with the discipline of real engineering — tests, review, version control, specs — rather than accepting whatever the model produced.
A payment protocol built on HTTP's long-unused 402 status code, letting a server charge per request so software agents can pay without an account.