​
Definition

An Agent is a durable member of an organization. Its record holds a role, capabilities, reporting line, runtime configuration, skills, budget, and permission state. An Agent Run is one bounded attempt by that Agent. The runtime is the program or service Rudder invokes for the attempt. Rudder coordinates the work, while the runtime performs it. The run keeps several kinds of evidence distinct: A readable summary never replaces the raw evidence. A partial or stopped stream also does not become a completed result merely because text exists.

​
One illustrative case

A release team has a writing Agent responsible for launch copy. The Agent keeps that responsibility after one announcement is finished. For a new announcement, Rudder starts one Agent Run through the configured local runtime. The run leaves a draft file, a short result summary, cost data, and a transcript that the reviewer can inspect. If the draft needs another attempt, the team starts another run. It is still the same Agent, but the two runs keep separate evidence and cost. Agent run detail

​
When an Agent is useful

Create an Agent when a repeated class of work needs a stable owner and operating frame. Use an existing Agent for a one-off request that fits its role. An Agent record is useful when later work should inherit the same responsibility, runtime, skills, budget, or escalation path. Runs can begin from Chat, Issue assignment, review routing, an Automation, or a manual or scheduled heartbeat. The trigger does not change the distinction: the Agent persists, the Run is bounded, and the runtime is invoked for that run.

​
Operating boundaries

Runtime capabilities vary. A local adapter may support managed tools, Browser control, session reuse, transcripts, or cost metadata that another adapter does not. Rudder must show those differences instead of implying provider parity. The highest-risk boundary is authority. A runtime tool can act only with the organization, Agent, Run, and permission context Rudder supplies. Tool availability does not bypass assignment, checkout, approval, budget, or workspace rules. For exact runtime choices, see Runtime types. To set one up, follow Configure an Agent runtime.