Terminology glossary

Status: Complete. Last reviewed 2026-08-28.

Selective one-sentence definitions for recurring interview vocabulary; follow each link for mechanism and consequences. Closely related terms are separated in common confusions, while authorities are indexed in the primary-source map.

Abstraction — A boundary that exposes a useful contract while hiding decisions callers should not depend on; see abstractions and their costs.

Agent — A model-driven loop that combines instructions, state, and tools to pursue an outcome; see tools, MCP, skills, and agents and tool/MCP/skill/agent.

Aggregate — A DDD consistency boundary whose root controls invariant-preserving access to the entities and values inside it; see entities, value objects, and aggregates.

Atomicity — The guarantee that a scoped operation’s changes commit or abort as one unit; see transactions and ACID and atomicity/idempotency.

Authentication — Establishing which actor or client is making a request; see Laravel authentication and authentication/authorization/isolation.

Authorization — Deciding whether an established actor may perform a specific operation on a specific resource; see gates and policies.

Backpressure — A mechanism by which a slower consumer bounds or signals upstream production instead of accepting unbounded work; see backpressure and bounded work.

Bounded context — An ownership boundary within which a domain term and model have one consistent meaning; see bounded contexts own meaning.

Cache consistency — The application policy defining acceptable freshness and repair for a derived cached copy; see the cache contract and cache/database consistency.

Cohesion — The degree to which responsibilities grouped in a module change for related reasons; see coupling and cohesion predict change cost.

Context window — The bounded sequence of input and generated tokens available to a model during one inference; see the probabilistic boundary and parameters/context/retrieval.

Coupling — The extent to which one component depends on another component’s existence, contract, timing, data, or implementation details; see coupling and cohesion.

Dependency injection — Supplying collaborators from outside an object instead of constructing or locating them during its behavior; see dependency direction, injection, and containers.

Dependency inversion — Structuring dependencies so high-level policy relies on abstractions rather than details it does not own; see dependency direction and dependency inversion/injection.

Embedding — A vector representation whose distance can be used as an imperfect proxy for semantic relatedness; see adaptation and retrieval.

Entity — A domain object defined primarily by identity and continuity rather than only its current attributes; see entities, value objects, and aggregates.

Evaluation — A repeatable measurement of an AI system against representative cases, explicit graders, and quality, safety, latency, or cost thresholds; see representative evaluation and deterministic tests/AI evals.

Idempotency — The property that repeating an operation has the same intended effect as performing it once; see API idempotency.

Invariant — A condition that must remain true across every allowed state transition and entry point; see validated input is still only input and validation/invariants.

Isolation level — The database contract governing which effects of concurrent transactions an operation may observe; see isolation and anomalies.

Job — An executable unit of work that often crosses a serialization, scheduling, retry, and process boundary; see the queued job lifecycle and event/command/job/notification.

MCP — The Model Context Protocol, a protocol for AI clients to discover and invoke server-provided tools, resources, and prompts; see tools, MCP, skills, and agents.

Metric — A numeric time-series signal aggregated over defined dimensions; see observability evidence and logs/metrics/traces.

Microtask — Browser event-loop work, including promise reactions, drained after the current task before the next task normally begins; see event-loop behavior and task/microtask.

OAuth 2.0 — A framework for delegated authorization to protected resources; see Laravel OAuth context and OAuth 2.0/OpenID Connect.

OpenID Connect — An identity layer over OAuth 2.0 that lets a client authenticate an end user and receive identity claims; see Laravel identity protocol context and OAuth 2.0/OpenID Connect.

Outbox — A table or durable record written with application state and later relayed to close the local transaction/publication dual-write gap; see transactional outbox.

Parallelism — Simultaneous execution on multiple processing resources; see execution and isolation models and concurrency/parallelism.

Policy — In Laravel, a class grouping authorization decisions around a resource or model type; see gates and policies.

Prompt injection — Untrusted content crafted to redirect a model or agent away from its intended instructions and controls; see injection and excessive agency.

Retrieval — Selecting external material at inference time for inclusion in model context without updating model weights; see adaptation and retrieval and parameters/context/retrieval.

Saga — A distributed workflow coordinating local transactions with durable state, retries, and compensating or forward-recovery actions; see sagas and compensations.

Service container — A registry and resolver used at composition boundaries to construct object graphs and manage dependency lifetimes; see Laravel container resolution and container/service locator.

Skill — A reusable instruction and resource bundle that teaches an agent a bounded workflow without itself being the action interface; see tools, MCP, skills, and agents and tool/MCP/skill/agent.

Task — A browser event-loop unit such as a timer or event callback after which queued microtasks are drained; see event-loop behavior and task/microtask.

Tenant isolation — The end-to-end invariant that one tenant cannot read, mutate, cache, deliver, or execute work under another tenant’s ownership; see tenant isolation.

Tool — An action interface a model or agent can call with structured input while the host retains validation, authorization, and execution responsibility; see AI tool boundaries.

Trace — A causally connected set of spans representing one request or workflow across boundaries; see observability evidence.

Transaction — A database unit of work whose guarantees include scoped atomicity and an isolation level rather than automatic multi-system consistency; see transactions and ACID.

Value object — A domain concept defined by its values, equality rules, and invariants rather than durable identity; see entities, value objects, and aggregates.

Variance — The rules determining how subtype relationships may change parameter and return types while preserving substitutability; see PHP variance.