Reference · updated September 25, 2026

AI-agent glossary

Plain-language definitions for reading an agent dashboard or a claim about a multi-agent system. Each term includes what not to infer from it.

Agent

Software that uses a model and a set of tools to take steps toward a goal. The tools and allowed actions define its reach. A registered agent is not necessarily working right now.

Swarm / multi-agent system

Several agents or agent roles coordinated around a goal. "Swarm" does not guarantee that they run simultaneously, cooperate well or produce correct answers. Look for a work trace and checks.

Orchestrator

The component that decides which worker gets which subtask, gathers results and decides what to do next. It may be a fixed program or an agent itself. Delegation adds coordination overhead, so a single agent can be better for a simple task.

Worker

An agent assigned a bounded piece of work. "Worker count" can mean created workers, available workers or actively executing workers; those are different figures.

Handoff

A transition from one agent or stage to another, with enough state for the next step to continue. In a state-driven workflow, the active step changes according to the task state. A handoff alone does not verify the output.

Evaluator / review gate

A check that compares a candidate output with a rule or evidence before accepting it. The reviewer can also be wrong. Stronger evaluation uses a reproducible test or an independent source, and records what was not checked.

Rate limit and quota

A provider's limit on requests or usage over a period. A fallback can try another permitted source when one is limited; it does not create free capacity or justify evading a provider's rules.

Cache

A stored result reused instead of running the same work again. It can save model calls, but stale or misapplied results can be worse than a fresh answer. Cache claims need a measured hit rate and quality check.

Throughput

How much work a system finishes in a stated time window, such as tasks per hour. It says nothing about usefulness unless "finished" has a clear definition and quality is measured too.

Proof record

A traceable account of a result: input, method, source or test, timestamp, supported conclusion and unresolved claims. A dashboard counter or a self-reported task completion is not the same thing as an independently checked outcome.