AI engineering glossary

A system that pursues a goal through a loop of observing, reasoning and acting until the task is complete, rather than answering once.

Retrieval where the agent explores a corpus iteratively with tools (search, browse, read, cross-reference) instead of a single top-k lookup.

Building a site so that AI assistants can retrieve, understand and cite it: crawlable, readable without JavaScript, structured, question-shaped, with visible authority and freshness.

Bounding how much work is in flight per worker, tenant and provider so that a growing queue slows intake instead of overwhelming downstream limits.

Asking the model to reason step by step before giving its answer.

The LangGraph component that persists graph state per thread, so a run resumes after a failure and can be rewound.

The memory file Claude Code reads at the start of a session: project conventions, architecture, commands and instructions, at project or folder level.

A database that stores each column separately, such as ClickHouse, which makes aggregation over large event tables fast.

A model acting on a desktop from screenshots, with its instructions translated into clicks, typing and key presses.

A design in which the product's behavior lives in typed configuration and the code interprets it, so a new customer is new configuration rather than new code.

A program that acts with its own authority on instructions supplied by someone with less authority; an agent executing injected instructions with your credentials.

The tokens the model can see in one request, budgeted across the system prompt, memory, conversation history and the response.

Watching someone do their real work in their own environment and asking about it as it happens, rather than asking them to recall it later.

Where a queue puts messages that failed every retry, so a person can inspect them.

Running a workflow as checkpointed steps so it resumes from the last completed step after a failure or restart.

Vectors that represent the meaning of text, compared by cosine similarity to find semantically similar passages.

Encrypting data with a data key and encrypting that key with a key held in a key management service.

A test of model output quality against defined criteria, run statistically because outputs vary between runs.

Processing many items in parallel and then collecting the results into one.

Giving the model examples of the input and output you want before the real input.

The model emits structured JSON that maps to React components from a defined catalog, rendered as it streams.

A formal description of a language, written in BNF as productions over terminals and non-terminals; parsers are built from it, and structured outputs are enforced by it.

Validation of an agent's input (reject harmful requests) or output (ensure quality) around the model call.

One agent transferring control to another, with the context the next agent needs to continue.

A shell command that Claude Code runs in response to an event, such as before a commit or after an edit.

Pausing an automated process for a person to confirm, approve or edit before it continues.

A caller-supplied identifier on a write so that a retried request produces the same result once.

Defining networks, databases, queues and services in files such as Terraform modules, so an environment can be planned, reviewed and recreated.

Using a model to score another model's output against criteria; cheaper than human evaluation and more flexible than programmatic checks.

A plain-text index at the root of a site listing its pages with one-line descriptions, written for AI assistants.

The Model Context Protocol: an open standard by which servers expose tools, resources and prompts to AI clients, including agentic coding tools.

A queue that tests each pull request against the latest main before merging it, so main stays green when many PRs land at once.

A program that reads a description of a language and produces a parser or interpreter for it; at product scale, a description of a product that generates the product.

One repository holding the apps and packages of a product, with one lockfile, one type-check and one CI.

A manager agent that delegates sub-problems to specializt agents and combines their results.

Reusing a cached prefix such as a long system prompt across requests, which model providers price at a large discount.

The abstraction, such as the Vercel AI SDK, through which an application calls any model provider with the same code.

Publishing events to a topic that several subscribers consume independently, so the producer is decoupled from the side effects.

Reason plus act: the agent alternates between reasoning about what to do next and calling a tool to do it.

A periodic full comparison of synced data against the source, to catch events that polling or webhooks missed.

An agent pattern in which the agent evaluates its own output and retries when the quality is low.

Postgres policies that filter every query by the caller's identity, so permission checks happen in the database rather than in application code.

Metric definitions kept in code as the one source of truth for revenue, active users and retention, which dashboards and agents query.

A custom slash command for Claude Code, defined as a Markdown prompt file, that automates a repeatable task.

Receiving the response token by token as it is generated, over Server-Sent Events, rather than waiting for the whole message.

Getting typed, validated data from the model, most reliably by defining a tool whose schema is the output you want.

A parallel coding-agent instance spawned for an independent task, whose result flows back to the parent agent.

The sampling control for randomness: 0 for deterministic extraction and classification, 0.7 and above for creative work.

The model requesting a function call with structured arguments, your code executing it, and the result returned to the model.

The record of one full request in an observability tool such as Langfuse, made of spans for each model call and operation.

One internal model per concept, such as contact or ticket, with mappings from each provider's shape and a raw field for the original payload.

A session in which a participant attempts realistic tasks with the product and thinks aloud while an observer records where they struggle, without helping.

A database for embeddings and similarity search: pgvector in production, ChromaDB for prototyping.

Presenting a corpus to an agent as files it can list, read and search, with each command translating to a database or vector query.

Detecting when a user starts and stops speaking, so the voice pipeline knows when to listen and when to respond.

A virtual private cloud: an isolated network in a cloud account, divided into subnets with security groups controlling traffic between tiers.

A second working directory for the same git repository, so several agents can work on different branches at once.