Task Center turns repetitive work into automated processes: a task runs step by step along a node graph, either when you start it or automatically at an interval or at a set time each day. Unlike a chat, a task's process is decided by the node graph you draw, and the Agent only handles the step that needs its own judgment, so results are more predictable. Select Tasks in the left navigation to open it.
Creating a task
- In Task Center select New task and fill in the task name and description.
- Select Edit node graph to open the canvas. The canvas starts with two nodes, Start and End.
- Drag the nodes you need from the node library onto the canvas and connect them in the order they run.
- Click each node and fill in its settings on the right. An Agent node, for example, needs a model, the task instructions, and a permission profile.
- Select Test run to check the process, and Save once it is right.
Available nodes
| Category | Nodes |
|---|---|
| Input and output | Start, End, Text input |
| Execution | Agent, LLM, Script, Code, HTTP request, Template transform, Variable aggregation |
| Control and notification | Condition, Iteration, Loop, Delay, Human confirmation, Notification, Remote delivery |
- Agent: starts one unattended Agent run. It has no chance to ask you anything while it runs, so the task instructions must state the goal, the input, the output, and the expected length.
- LLM: one call to a model without tools, suited to classification, extraction, rewriting, and summarization, and faster and more predictable than an Agent node.
- Human confirmation: the run pauses here and you select Approve or Reject. A task that contains a Human confirmation node can only run on the canvas, never on a schedule.
- Notification: sends a system notification; a failed send does not stop the task.
- Remote delivery: sends a piece of text to a Remote Control channel you have connected; first connect the account and allow the chat under Remote Control, as described in Remote control and channels.
Building with AI
If you would rather not drag nodes yourself, select Build with AI on the canvas and describe what you need in plain words. The task-building Agent changes the nodes and connections on the canvas directly, and can also run the process to test and repair it.
Building a task from a chat
- Choose Task in the mode selector above the input box.
- Describe the process you want to automate in plain words, and the Agent drafts the task in the chat.
- When you are happy, select Save to Task Center.
- Go to Task Center to finish setting permissions, the schedule, and the remote account.
If Task Center has changed the task after you saved it, the older draft in the chat can no longer overwrite it; carry on editing in Task Center.
Running tasks and reading history
- Running by hand: select Run now in the task list.
- Test run: select Test run on the canvas, which saves the current canvas first and then runs once. The Run panel on the right shows each step, the logs, and the variables.
- Run history: open Run history to see the status of each node, the output, and the files produced in every run. Each task keeps its latest 100 records, and test runs count too.
If any node fails, the whole run stops at once and is not retried automatically. The branch of a condition that was not taken shows Skipped, which is normal.
Running on a schedule
- Open the task settings and choose By interval (every N minutes) or Daily at a set time, and set the time. Choose No schedule (run manually) and the task can only be started by hand.
- Choose an Unattended permission profile for the Agent node, and fill in Tool allowlist (required for scheduled tasks).
- Fill in the working directory for a Script node (required for scheduled tasks).
- Tick Enable schedule and save. In the task list you can Enable schedule or Pause schedule at any time.
Unattended permission profiles
| Profile | What the Agent can do |
|---|---|
| Read-only (read-only) | Read files, search, read memory, research online |
| Workspace read-write (workspace-write) | Everything read-only allows, plus creating and changing files; commands and the browser are off by default |
| Full (full, including shell and browser) | Everything, including running commands and browser automation |
On a schedule no one is there to select confirm, so tools outside the allowlist are refused outright. Grant the least permission you need: Agent nodes are read-only by default, upgrade only when files really have to change, and use the Full profile only for tasks you fully trust.
Things to keep in mind
- Schedules run in the background of the desktop app. They keep running when you close the window or minimize to the tray, but no scheduled task runs after you quit the app.
- A run missed while the computer slept or the app was closed is not caught up; it is recorded as Skipped.
- A task can have only one run at a time; if the previous run has not finished, the run due now is skipped. Set the interval longer than a run takes.
- For a critical task, add a Notification or Remote delivery node at the end so you hear about a failure in time.