​
Completed state

At the end of this guide, the organization has one new active Agent with a recognizable responsibility, valid reporting line, bounded capabilities, saved runtime configuration, and one inspectable low-risk run. Create a durable Agent for repeated ownership. A one-off prompt that fits an existing role does not need a new Agent.

​
Prerequisites

You need an organization, permission to create Agents, a repeated responsibility that is not already owned, and a runtime that Rudder can test. This example adds an Agent responsible for launch copy and uses a small announcement as the first task.

​
Create the launch writer

  1. Open Agents and choose New Agent.
  2. Name the Agent Launch Writer and give it a title that describes its position, such as Release Communications Writer.
  3. Choose the role used by the organization structure and set Reports to to an active Agent in the same organization.
  4. Select and configure the runtime, then run Test runtime chain.
  5. For a local runtime, write the prompt template that frames the role and its boundaries. Set the run policy, and enable only the optional Skills this responsibility needs.
  6. Choose Create agent. A policy may leave the new Agent pending approval instead of immediately invokable.
  7. After creating the Agent, open Configuration. Write Capabilities as an acceptance boundary, then review Permissions on the same page.
  8. Open Budget and set the monthly limit appropriate for this role.
  9. Assign a low-risk launch-announcement Issue with clear acceptance criteria, or ask for the same bounded draft in Chat.
  10. Inspect the Agent Run, draft output, validation note, cost when available, and any blocker.

​
Success signal

Agent detail shows the intended role, reporting line, capabilities, permissions, runtime, Skills, and budget. The low-risk run uses that runtime and leaves an output plus a clear completion or blocker signal on its source Chat or Issue.

​
Recovery

If creation is blocked, check organization scope, create-Agent permission, reporting-line eligibility, and runtime test output. Pending-approval, paused, terminated, and revoked-access Agents are not ordinary invokable Agents. If the first run accepts the wrong work, tighten the capability statement and assignment context. If it cannot start, repair the runtime with Configure an Agent runtime, then repeat the low-risk run before assigning broader work.