Agents glossary

Terms used in the Agents API and the agent pages. Shorter definitions of the core API terms are on the main Glossary.

Agent definition

A saved configuration of model, instructions, tools, skills, MCP servers and limits. Changing it creates a new version. See managed agents.

Agent version

An immutable snapshot of a definition. Sessions record the version they ran, so behaviour can be traced back. See versions.

Approval

A pause before a listed tool runs, until a person sends user.approval with approve or reject. See permissions.

Consolidation

A background job that merges duplicate memories, settles contradictions and expires stale ones. See memory consolidation.

Delegate

Another agent that the primary agent may hand a sub-task to. It appears to the model as a tool. See threads.

Deployment

A rule that starts sessions on a schedule or from an inbound webhook. See deployments.

Deployment run

One firing of a deployment, with the session it created.

Environment

A container definition (image, packages, network rules, limits) where an agent’s tools run. See environments.

Event

One entry in a session’s history, such as message.completed or user.tool_result. See events.

Executor

Who runs a tool call: syntic (server tools, environment tools, MCP) or client (your code answers with user.tool_result).

Lease

A runner’s temporary claim on a work item, 60 seconds long, extended by heartbeats. An expired lease returns the item to the queue.

Memory scope

A label such as customer:4821 that partitions a store. A session reads and writes only its scope plus global.

Memory store

Persistent notes an agent can read and write across sessions, organised by scope. See memory.

Outcome

A short result the agent records when a session’s goal is reached, such as resolved or escalated.

Overlap policy

What a deployment does when the previous run is still going: skip, queue or allow.

Self-hosted runner

A machine you operate that claims and runs tool calls for an environment with hosting: self_hosted. See self hosted runners.

Session

One long-running conversation with an agent, held on the server with its full event history. See sessions.

Session budget

Caps on turns, tokens or dollars for one session. Reaching one ends the turn with an error. See budgets.

Session resource

A file, Git repository or memory store mounted into a session. See session resources.

Skill

A packaged folder of instructions and files the model loads only when a task needs it. See skills.

Thread

A line of work inside a session. The primary thread is the main conversation; delegate threads run sub-agents. See threads.

Turn

One round of work: from a user message (or trigger) until the agent stops and waits. A turn can include many model calls and tool calls.

User profile

Facts about one end user (name, locale, time zone, attributes) given to every session started for them. See user profiles.

Vault

Write-only storage for secrets that tools and MCP servers use. Values never reach the model. See vaults.