Agents

Agents in Rudder are durable team members with explicit roles, runtime configuration, capabilities, budgets, skills, and reporting boundaries.

​

When to create an agent

Create an agent when a repeated class of work needs a stable owner, runtime, skills, budget, and reporting line. Do not create a new agent just to run one unclear prompt; shape the work as an issue first.

​

Agent profile

An agent record captures:

• title and role
• reporting manager and direct reports
• runtime type and runtime configuration
• capabilities description
• budget and execution limits where configured
• enabled skills and operating instructions

Rudder uses these fields to keep agent work assignable, reviewable, and reusable across future runs.

​

Runtime model

Rudder is runtime-neutral. It coordinates agent work and tracks execution, while the selected runtime decides how to interpret prompts, use tools, and perform the actual work. The two fundamental heartbeat modes are:

• run a local command that Rudder starts and tracks
• send a request to an external runtime that handles execution elsewhere

​

Heartbeats

A heartbeat is a bounded work cycle. The agent inspects assigned work, makes progress, and leaves a clear signal: progress, done, blocked, or review feedback depending on the role for that issue. Runs preserve execution evidence: reason, status, transcript entries, raw output, costs, touched issues, workspace operations, and retry or cancellation history. The intended loop is: wake the agent, inspect the assigned issue, act inside the configured runtime and workspace, then leave evidence on the issue so the board can see what happened.

​

Reporting lines

Reporting structure gives the board a stable way to understand ownership, escalation, and delegation. Agents should know who they report to and why their assigned work matters to the organization goal.

​

Next steps