Glossary
The vocabulary of AI-written enterprise software.
Definitions for the terms this site uses, written to be quoted. Each entry opens with one self-contained sentence, then explains what the term means in practice and links to the articles and pages that use it.
22 terms
-
Agent guardrails
Agent guardrails are the constraints placed on what an AI agent may do — from prompt instructions and output filters to runtime permission checks, approval gates, and audit logging — and they divide into the ones an agent can ignore and the ones it cannot.
-
AI agent runtime
An AI agent runtime is the execution environment that runs an agent loop — calling the model, invoking tools, carrying state and memory between steps, and handling retries, timeouts, and concurrency — so that an agent runs as a supervised program rather than as a single prompt.
-
Application definition vs. runtime
Application definition vs. runtime is the architectural split between what an application is — a typed, version-controlled declaration of its objects, fields, relationships, permissions, workflows, actions, and APIs — and what runs it: an engine that reads that declaration on every request and derives the database, the API, the interface, and the permission and audit enforcement from it, so that the definition stays portable and the engine stays replaceable.
-
Application metadata
Application metadata is the structured, typed definition of a business application — its objects, fields, relationships, views, permissions, workflows, actions, and APIs — stored as source files that a runtime reads and executes directly, rather than as generated application code.
-
Approval workflow
An approval workflow is a business process in which a proposed change is held in a pending state until a person with the authority to decide records an explicit approval or rejection, so the change takes effect only after someone accountable has signed for it.
-
Audit trail
An audit trail is a chronological, append-only record of who did what to which data and when, written by the system that executed the actions rather than by the parties performing them, so that history can be reconstructed and attributed after the fact.
-
Comprehension debt
Comprehension debt is the widening gap between the software an organization runs and the software anyone in it can still explain — a liability that accrues whenever working code ships faster than the understanding of it, and that comes due at the first change nobody dares to make.
-
Declared vs. enforced
Declared vs. enforced names the gap between what a system’s configuration lets you write down and what its runtime actually checks at execution time — a declaration that nothing enforces reads like a guarantee and behaves like a comment.
-
Definition layer
The definition layer is the tier of a software stack that holds a business’s own declarations — which objects exist, how they relate, who may do what to them, and which steps require approval — as versioned files the organization owns, sitting above the runtime layer that executes those declarations and outside any single vendor that happens to ship an engine for them.
-
Governed runtime
A governed runtime is an application runtime that enforces identity, permissions, approvals, and audit logging on every read, write, and tool call, so that an AI agent and the person it acts for are held to exactly the same rules.
-
Governed tool layer
A governed tool layer is the layer between an AI agent and a business system in which the agent calls controlled actions rather than raw interfaces, so that every call carries the initiating user identity, is checked against that user permissions, pauses for approval when the action requires one, and lands in an audit ledger.
-
Knowledge graph
A knowledge graph represents an organization's information as nodes and typed edges — the entities it cares about and the labelled relationships between them — so that questions about how things connect can be answered by traversing those relationships rather than by joining tables.
-
MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard that defines how an AI application connects to external tools, data sources, and prompts over a uniform client–server interface, so that any compliant client can use any compliant server without integration code written for that specific pair.
-
Metadata-driven development
Metadata-driven development builds an application by declaring its objects, fields, permissions, workflows, actions, and APIs as typed metadata that a runtime executes directly, instead of writing the code that implements them — a distinction that becomes decisive once an AI agent is doing the writing, because the declaration is small enough for a human to review and the runtime, not the generated code, is what enforces the rules.
-
Object type / action type
Object type and action type are the two halves of an ontology's vocabulary: an object type declares a class of business entity together with its properties and its links to other types, while an action type declares a permitted, permission-checked operation that changes those objects — so that writing is defined as explicitly as reading.
-
Ontology
An ontology, in the enterprise sense, is a formal model of a business domain — the classes of things the business deals with, the properties that describe them, and the relationships that connect them — expressed in a machine-readable form so that software, and now AI agents, can reason about the business rather than about raw tables.
-
Permission model
A permission model is the structure a system uses to decide who may do what — which principals exist, which containers hold grantable capability, how multiple grants combine, and what happens when nothing grants access — as distinct from any individual permission expressed within it.
-
Reviewable diff
A reviewable diff is a change small enough, and expressed at a high enough level, that the person accountable for it can read every line and judge on business terms whether it is right — the property that decides whether AI-written software gets merged with a signature behind it or on faith in green CI.
-
Row-level security
Row-level security (RLS) is an access-control mechanism that restricts which rows of a table a given user may read or write by attaching a predicate to the table itself, so the restriction is applied inside the query rather than by the application code that issues it.
-
Semantic layer
A semantic layer is a governed set of business definitions — metrics, dimensions, and entity names — placed between raw data storage and the tools that query it, so that "active customer" or "net revenue" resolves to one agreed calculation no matter who asks or which tool they ask from.
-
Typed metadata
Typed metadata is application metadata whose keys and values are constrained by a published schema, so that an unknown field type, a misspelled permission, or a workflow step pointing at an object that does not exist is rejected at a validation gate when the definition is written, instead of being discovered in production after it runs.
-
Vibe coding
Vibe coding is the practice of building software by describing what you want to an AI model in natural language and accepting the code it produces without reading or understanding it — a term Andrej Karpathy coined in February 2025 for giving in to the vibes until you "forget that the code even exists."