Skip to main content
The Lava desktop agent (the agent) asks before high-stakes actions. By default (Normal mode), the agent works through a task on its own and pauses for your approval only before it sends a message or email, deletes data, makes a payment, or shares something externally. Routine reads and low-risk writes run without a pause. You stay in control. This page is about how to use that well. You can adjust this from the Mode dropdown in the composer: Careful pauses before every write; Auto runs the entire task without pausing. See How Lava Desktop Works for the full approval model.

​
You Remain Responsible for What You Approve

The Approve action card shows you a plain-language summary of exactly what the agent is about to do. Read it before you click Approve. The approval step is your check, not a rubber stamp. The agent’s description of the action comes from its own understanding of the task. That understanding can be wrong. When something in the card looks unexpected (the wrong recipient, an unintended scope, an action you didn’t ask for), decline it and redirect the agent with a more specific instruction. Approving an action means you take responsibility for it. Treat each approval card the same way you would treat an unsent email: read before you send.

​
Start Small and Build Trust

Start with low-stakes tasks in apps where a mistake is easy to reverse. A misread, a draft that goes nowhere, a calendar event you can delete: these are good places to learn how the agent interprets your instructions before you hand it higher-stakes work. Once you have a feel for how it handles your phrasing and which apps it navigates well, you can delegate more. Trust is built by observation, not by assumption. Good starting tasks:
  • Summarize recent emails or messages (read-only, no approval needed)
  • Draft a reply without sending it
  • Create a calendar event you can review before the agent finalizes it
Work up from there to tasks that send, post, or modify.

​
What to Be Cautious With

Some work carries more risk than others. Use extra care (or handle it yourself) when:
  • The action is hard to undo. Deleted records, sent bulk messages, and removed calendar entries are difficult or impossible to reverse. Decline and do these by hand if the stakes are high.
  • The scope is large. A task that touches many records or many people in one run has a wider blast radius if something is misunderstood. Break it into smaller pieces first.
  • The data is sensitive. The agent reads the active tab for context. If a tab contains confidential information you would not want processed externally, either close that tab or keep the agent out of that app for the session.
  • The instruction is ambiguous. The agent asks you when it cannot ground a reference, but it may not catch every ambiguity. If you are not sure how the agent will interpret a task, ask it to describe what it plans to do before you let it proceed.

​
The Agent Can Make Mistakes

The agent can misread context, misinterpret an instruction, or take the wrong action on the right intent. The approval gate reduces the risk of a bad outcome, but it does not eliminate it: you can approve an action that turns out to be wrong. When a task goes differently than you expected:
  • Decline subsequent approval cards and let the agent stop.
  • Correct the result manually where needed.
  • Rephrase the task and try again with a narrower scope.

​
How Your Data Is Handled

The agent reads the active tab for context and acts through the Lava gateway on your behalf. For details on what data Lava collects, retains, and how it is used, see the Lava Privacy Policy.

​
Next Steps

Working with the Agent

Give the agent tasks, follow its progress, and use the approval and stop controls.

How Lava Desktop Works

Understand the gateway, why high-stakes actions pause for approval, and how the agent grounds its actions.