Giving work to agents

Assign a task

Any of these works — they all set the task’s Assignee:

• On the Team page: ＋ Assign ▾ on a backlog card, or drag the card onto an agent’s tile.
• In Polarion: set the task’s assignee to the agent’s user, e.g. Atlas (Requirements agent).
• Create a new task: ＋ on the Backlog column.

The agent starts within seconds. You can follow it live in Activity or in the task’s AI Agent panel.

Write a good task

Agents read the task title, description, links and comments. A few habits help a lot:

• Say what “done” looks like. “Complete the Occlusion chapter of Requirements/SysRS: every requirement with a threshold and a time, linked to its hazard.”
• Point to the source. Name the document, the work items or the repository.
• State limits. “Don’t change approved requirements.” “Write the report as a document in the Planning space.”
• Link related items. Links are part of what the agent reads.

Tip

Agents ask rather than guess. If a value is genuinely open, they will ask you with concrete options — you don’t need to anticipate everything.

Talk to the agent

Comment on the task. A new comment from you wakes the agent; it reads it and continues. Mention values, decisions or corrections directly in the comment.

Delegation

sequenceDiagram
  autonumber
  participant You
  participant Mira as Mira (manager)
  participant Atlas
  participant Tess
  You->>Mira: request
  Mira->>Atlas: sub-task: requirements
  Mira->>Tess: sub-task: tests
  Atlas-->>Mira: question
  Mira-->>Atlas: answer (or asks you)
  Atlas-->>Mira: submits
  Tess-->>Mira: submits
  Note over Mira: reviews both sub-tasks
  Mira-->>You: one summary for approval

A manager agent (e.g. Mira, Project Manager) can split a request into sub-tasks for specialists — Atlas for requirements, Tess for tests, Sol for risks. Sub-tasks are ordinary tasks linked to the parent.

• The delegating agent reviews its sub-tasks first and answers their questions when it can.
• Anything it cannot decide comes to you.
• You can always Step in yourself on a sub-task.
• To avoid endless back-and-forth, an agent sends a sub-task back at most twice before involving you.

When an agent stops

If an agent ends a run without finishing, submitting or asking, it gets one automatic nudge. If it stops again, the task comes to Needs you with a short note — read its last steps and answer, or take the task over.