Glossary

Agent terminology is used loosely across the industry. These are the meanings this documentation commits to.

Agent

A system that pursues an objective by planning, acting through tools, observing results and adapting.

Artifact

A durable output of a run, carrying provenance and retention policy.

Approval

A human decision required before a declared effect may occur.

Budget

A hard ceiling on steps, time, tool calls or spend for a run.

Capability

A named grant that makes a class of tools available.

Checkpoint

A durable snapshot of run state enabling resume or fork.

Connector

A packaged integration exposing an external system as typed tools.

Effect class

The reversibility and visibility category of an action, used to decide gating.

Grader

A programmatic or rubric-based check that scores an evaluation case.

Objective

The outcome a task must achieve, stated as a result rather than a procedure.

Observation

A recorded result of an action that can revise the plan.

Policy engine

The component that evaluates every dispatch against workspace rules, outside model context.

Provenance

The record of where a fact or artifact came from and how it was produced.

Run

One execution attempt of a task.

Sandbox

An isolated, disposable environment where code executes.

Simulation

Executing a transaction against forked state to decode its effects before submission.

Step

One planned unit of work within a run.

Task

A unit of work defined by an objective, constraints and completion criteria.

Tool

A typed, executable capability the agent can invoke.

Trace

The full structured record of a run: plan, calls, observations, decisions.

Untrusted content

Data from outside the trust boundary; may inform reasoning, never grants authority.