An Agentic Task lets you interact with an AI agent while a task is running in a Cutover runbook. You can use the agent to investigate issues, gather information, understand the state of a system, draft commands, or perform approved actions. The conversation is part of the task, so the messages, responses, approvals, and outcomes remain available to the team as part of the task's execution history.
Use the AI chat panel to communicate with the agent while the task is in progress. You can:
When the agent is generating a response, a streaming indicator appears on the task in the task list.
You can send messages when:
If messaging isn't available, Cutover displays a message explaining why.
The agent's capabilities depend on the type of agent attached to the task.
Cutover AI Assistant automatically has the context of the runbook and task.
You can ask it about the runbook, retrieve runbook information, investigate issues with the runbook’s execution, and perform supported runbook actions.
The assistant uses your Cutover permissions, so it can't access or change anything that you don't have permission to access or change.
A custom AI agent is provided and configured by your organisation.
What the agent can see and do depends on how it has been configured and integrated with Cutover. Custom agents don't automatically receive the full runbook and task context. This information can be provided to the agent through the integration, for example by instrumenting the agent with Cutover MCP.
For information about configuring AI agents, see Develop and Configure AI Agents.
An agent may ask you to approve an action before it can be performed. For example, the agent may propose running a command or, when using Cutover AI Assistant, making a change to the runbook.
When approval is required, an approval card appears in the conversation.
You can choose:
Allows the agent to perform the proposed action once. The agent then continues its response using the result of the action.
Allows the proposed action and automatically approves future matching actions within the same conversation. You can choose to allow the specific action or all actions for the conversation.
Automatically approved actions are labelled Auto-approved in the conversation. These approvals are associated with the task conversation, so they apply to the conversation rather than only to the browser or user who granted the approval.
Rejects the proposed action. The action isn't performed and the agent is informed that the request was declined.
Approval and rejection decisions are recorded in the runbook audit log.
This provides a record of who made the decision, the action involved, the conversation, and the decision made.
The action's arguments and execution result aren't recorded in the audit log. The result remains available in the conversation history.
Certain actions, such as deleting a runbook task or deleting a stream, are considered high impact and will always require approval, even if you chose Approve all tools for the conversation.
The conversation belongs to the task, not to an individual user. Anyone who can view the task can view its conversation history. If multiple users have the task open, they can see messages and agent responses as they happen.
This means team members can take over or follow an AI-assisted investigation without losing the context of the conversation.
Your permissions determine what you can do with the Agentic Task.
Users with task view access can:
Users with task edit access can:
The agent also operates within the permissions of the user interacting with it.
If the run is paused while the agent is responding, Cutover warns you that the agent is still responding.
Pausing ends the current response. Any content generated before the pause is retained.
The conversation history is preserved. When the run resumes and the task is in progress again, you can continue the conversation.
When the task is completed, skipped, or abandoned, the chat panel closes. The conversation history remains available to users who can view the task.
If the run is cancelled, the conversation also closes. Any pending tool requests are disabled after the chat panel is closed.
By default, agents aren't available during rehearsal runs. If the agent has been configured to run during rehearsals, you can use it in the same way as during a live run.
The task displays a status to show whether the agent is available.
The task is connected to the configured agent and ready for interaction.
The agent couldn't be reached or the connection encountered an error.
The conversation/chat panel has closed because the task was completed, skipped, abandoned, or the run was cancelled.
The agent couldn't be reached or the connection encountered an error. If the underlying issue has been resolved, send another message to try again.
If the problem continues, contact the person responsible for the agent or integration.
Agent responses have a time limit. If a response takes too long, it stops and an error is displayed in the conversation.
Once the issue has been resolved, send another message to continue the conversation.
If the message controls are unavailable, check that:
