My Tasks¶

Tasks are the primary entry point for your collaboration with the Agent. Compared to casual questions, tasks are better suited for carrying work that has goals, context, a process, and results.

Create a Task¶

Click New Task in the Agent workspace and enter the request and required context directly. After you send the first message, the Agent generates a task name from the request and begins execution.

State the target, scope, time range, expected output, and important restrictions when they are known. This gives the Agent a clearer starting point than a general request such as "take a look."

Selecting New Task opens an empty draft. The task is created and appears under My Tasks only after you send the first message.

Choose Permission Mode and Model Intelligence¶

Before sending a message, adjust the composer options for the task's risk and complexity:

Permission mode and Model Intelligence can also be changed after creation. Subsequent messages use the current selection.

Find and Manage Tasks¶

My Tasks separates tasks by status:

• Active tasks can still receive messages, run, or wait for processing.
• Completed tasks have been formally closed. They are view-only and cannot receive new messages or attachments.

The task creator can edit its title and description so other participants can identify its goal and context. The creator can also delete it. Deleting an active task stops execution immediately and removes it from every participant's task list; deleting a completed task also removes it from every participant's list. Deletion cannot be undone, so retain any required conclusions and attachments first.

When the Agent is being created, offline, or suspended, tasks are view-only: messages cannot be sent and titles or descriptions cannot be edited. The task creator can still use the delete action shown on the page. Other participants can perform only the actions available to their current permissions.

Advance the Session¶

After entering the task session, you can tell the Agent what to do next, just like chatting.

Using @ in the session input lets you reference resources in the current workspace so the Agent can analyze a specific object. Supported resources include services, applications, dashboards, hosts, containers, and knowledge bases.

After referencing an entity, the Agent can narrow the analysis scope by combining the entity's context, for example:

• Referencing a service to analyze abnormal requests, error rates, latency changes, and upstream/downstream impacts of the service.
• Referencing an application to analyze access, performance, errors, and user experience issues related to the application.
• Referencing a dashboard to interpret metric trends, anomaly points, and correlated data within the dashboard.
• Referencing a host to check host resources, load, logs, and related alert clues.
• Referencing a container to analyze container status, resource usage, restarts, logs, and the cluster environment.
• Referencing a knowledge base to answer questions, verify procedures, or organize recommendations using existing Runbooks, operating standards, or team knowledge.

After entering @, find and select the resource by type. The resources shown depend on the current workspace data and your access permissions. When referencing a knowledge base, state how the Agent should use it, for example, "Use this Runbook to check the current alert." A reference provides task context and does not automatically execute operations described in the content.

In the session, you can also:

• Use Skills to have the Agent process in a specific way.
• View the Agent's thought process and execution results.
• Follow subagent or external Agent collaboration.
• Stop the Agent during execution.
• Initiate new requests when needed.

Task Queue¶

If the Agent is already running, a new request is added to the Task queue and processed in order. You can review the running and waiting items and remove a queued task before it starts.

Queue instructions that depend on a known earlier step. If the next direction depends on the current result, wait for that run to finish before adding it so an older queued instruction does not conflict with the latest conclusion.

If the task is complex, it is recommended to let the Agent formulate a plan first, then decide whether to execute it.

Plan Mode¶

Plan mode is suitable for complex tasks, such as problem troubleshooting, solution design, release checks, refactoring assessments, etc.

When using plan mode, the Agent will first help you clarify goals, scope, steps, and risks before proceeding to execution. This reduces misunderstandings and prevents the Agent from taking direct action with insufficient information.

It is recommended to use plan mode in the following scenarios:

• The task affects multiple people or systems.
• It involves production environments, releases, permissions, or external services.
• Multi-step execution is required.
• Scope and success criteria need to be confirmed first.
• You want the Agent to propose a solution before deciding whether to implement it.

Complex Tasks and Subagents¶

For a complex task with independent work items, the Agent can create temporary subagents in the current task. They can perform research, analysis, validation, or content generation in parallel, and the main Agent consolidates their results. No separate configuration is required.

For example, ask the Agent to:

• Analyze several independent services or data objects;
• Review code, tests, and documentation separately;
• Validate an issue against different evidence sources;
• Produce independent materials and combine them into a final report.

The Subagent runs card under a task message shows each subtask's status, progress, result, or error. Statuses can include Starting, Running, Retrying, Possibly stalled, Canceling, Review required, Timed out, Failed, Canceled, Unrecoverable, and Completed.

Subagents use the current Agent's existing tools, permission mode, and safety boundaries; splitting work does not grant additional access. If one subtask fails, other independent work can continue. The main Agent summarizes completed, failed, and remaining items from the available results.

A subagent cannot create another level of subagents. For deeper decomposition, ask the main Agent to plan the work items directly and consolidate them.

To improve task decomposition, state:

• Which work items are independent;
• The input, scope, and expected result for each item;
• How the final results should be consolidated, compared, or prioritized;
• Which steps have strict dependencies or write to the same resource.

Steps with ordering dependencies, multiple writes to the same object, or sequential approvals should remain sequential instead of being forced into parallel execution.

When the task needs the specialized capability or data scope of another deployed Agent rather than a temporary subagent, configure Agent Collaboration (A2A) first.

Task Insights¶

Task Insights organizes key information from the session to help you quickly understand what has happened in the current task.

Viewable information includes:

• Evidence.
• Hypotheses.
• Analysis.
• Anomalies.
• Impact scope.
• Conclusions.
• Action recommendations.

Task Insights supports timeline and list views. The timeline is suitable for reviewing the task progression process, while the list is suitable for viewing key information by type. If there are many task messages, you can first check Task Insights to grasp the key points.

Task Attachments¶

Task Attachments collect files associated with the current task. You can upload images or files from the message composer as context for the Agent. The Agent may also produce PDFs, spreadsheets, reports, exports, and other attachments during analysis or result generation.

In the Task Attachments page, you can:

• View files uploaded by users and generated by the Agent.
• Download corresponding files as needed.

Remove unnecessary sensitive information before uploading and confirm that the Agent and task participants are allowed to view the content. Completed tasks cannot receive new attachments.

Approval¶

When the Agent needs to perform high-risk actions, approval requests may appear. You can view pending approval items in the session and decide whether to allow them.

Before approving, it is recommended to confirm:

• What action the Agent intends to perform.
• Which systems or data this action affects.
• Whether there are rollback or recovery methods.
• Whether this step is truly necessary for the current task.
• Whether the operation complies with team processes.

Do not skip approval judgment because the task is urgent.

Stop Execution¶

If you find the Agent is heading in the wrong direction, the task scope needs adjustment, or the current execution is no longer needed, you can click Stop Execution. After stopping, you can add new instructions and then re-initiate the request.

Stopping execution only interrupts the currently in-progress response or operation; it does not mean the task is completed. The task can still continue to advance, or you can let the Agent execute again after adding instructions.

Complete a Task¶

Completing a task formally closes the work. When the objectives have been achieved or the task no longer needs to continue, click Complete Task, then confirm Complete and Generate Report.

The Agent organizes the current session and generates a task report before the task is marked Completed. This is a one-time action: a completed task cannot be restored, and no new messages, attachments, or requests can be added. If you only need to interrupt the current run and plan to continue later, use Stop Execution instead.

After a task is completed, the team can review key conclusions, main evidence, Task Insights, and output attachments around this task, facilitating review, handover, or subsequent tracking.

Before completing, it is recommended to confirm:

• Whether the goals have been achieved.
• Whether the key conclusions are clear.
• Whether the attachments that need to be retained have been generated and are downloadable.
• Whether the main evidence and conclusions are visible in Task Insights.
• Whether subsequent action items have been clarified.