Working with the agent

Give the agent a goal, useful context and the boundaries of the task. It can compose files, tools, widgets and connected apps; you do not need to design the implementation in advance.

Build a useful dashboard

Start with a form, a list and persistent data. Add filters and automation after the first version works.

Build a service-request tracker here: customer, subject, status and next action. Let me add and edit requests. Persist the data and explain how to back it up.

Turn a message into work

Connect Email first. The agent can read the selected message, create a task and prepare a reply in the composer. Inspect dates and recipients before asking it to send.

Create a task from this email, due Friday at 10 in my timezone. Open a reply draft confirming that deadline; leave it ready for my review.

Ask from your documents

Configure embeddings, upload a short source and give its collection a clear description. Ask for citations and test a question the document cannot answer.

Using the equipment manual collection, explain the maintenance interval and cite the source pages. Tell me if the document does not specify it.

Schedule a recurring result

Ask for the timing, timezone, intended output and delivery channel. An agent-triggering schedule can use the model; a data-refresh script can run without an AI request.

Every weekday at 8 in my timezone, summarize overdue tasks and send the summary to my linked Telegram chat. Show me the schedule after creating it.

Work together in an open form

The editor tool includes the open note, mail draft, task or event and its revision. Notes save immediately; task/event forms and email sending need their normal confirmation. Save important browser drafts before closing their tab.

Rewrite the open note into a short meeting report, preserving the decisions and action owners.

Keep working when you leave

Dashboard tasks continue when you open Settings, change dashboards or close the browser. Return to the same conversation for progress and results. Board indicators show running tasks and tasks waiting for input. Stop cancels the task and its helpers; closing a tab does not. Pending approvals and secure secret requests reappear on reconnect.

A server restart interrupts running tasks; saved history remains, but operations are not automatically replayed. Inspect what completed before asking to continue. Normal limits are one active task per conversation, four per user, sixteen per server and two hours per task including approval waits. Without a database, recent results are retained only in memory. Hanging up a phone call ends its voice agent.

Independent work with subagents

For substantial independent work the agent can delegate to helpers, then integrate their results. They use the same model, execution workspace, permissions and model-call budget. Give helpers distinct files to avoid conflicting edits. They cannot delegate again, request credentials or control your open editor. Up to three run concurrently, six per task, for at most fifteen minutes each. Unfinished helpers stop with the parent task.

Approve actions and understand limits

Approval: manual pauses each dashboard tool call for Approve or Reject; auto executes immediately. This setting is remembered per browser and fixed for the active task. External channels, cron and webhooks run automatically because no approval interface is present. Shared-mode tool policies still apply. Max iterations is a model/tool-loop budget, not a monetary cap on API usage or code run by tools.

Restore a file after a bad change

Ask the agent to inspect workspace_history and restore one path from the selected commit. The restoration is itself recorded and reversible. Workspace history is not a backup of PostgreSQL, external accounts or Docker volumes; files excluded from versioning are not covered.