Working with the Agent

The Lava desktop agent (the agent) works across every app you have connected. You give it a task in plain language; it reads context from the active tab, acts through your connected apps, and brings the result back to you. This page covers how to give it tasks, follow its progress, and control what it does.

​

The Agent Panel

The agent panel lives on the right side of the window. Press Cmd+\ to show or hide it. When you haven’t given the agent a task yet, the panel shows:

• “Ask your agent anything”: type any request in the input bar at the bottom, or press Enter to send.
• Suggestion buttons: a few ready-to-use prompts based on your connected apps. Click one to send it directly.
• “See your previous conversations”: link to your conversation history.

The input bar placeholder reads Ask the agent - Enter to send. Type '/' to trigger a workflow. Type / to browse saved workflows, or just type a task directly.

​

Give the Agent a Task

Type your request in the input bar and press Enter. A few principles make tasks work better:

• Be specific about what you want. “Summarize unread emails from this week, grouped by sender” works better than “check my email.”
• Name the app when it matters. “In Slack, find the thread where we discussed the launch plan” focuses the agent on the right place.
• Start with one step. The agent handles multi-step tasks, but short first requests let you see how it interprets your intent before it acts broadly.

​

Run a Workflow

Type / in the input bar to open the workflow picker. Workflows are saved task recipes, a single trigger that runs a named sequence. Select one to start it.

​

Cross-App Tasks

One task can span multiple apps. You might ask the agent to summarize a thread in Slack and then draft a follow-up in Gmail. The agent handles both without you switching between them. As it works, the agent reads the active tab for context, then reaches through the Lava gateway to act in the relevant apps. It does not fabricate context from one app when it is working in another.

​

Follow the Agent’s Work

While the agent is running, the panel shows its progress as it goes. When it acts in a specific app, it surfaces a “Show in [app]” link next to that step. Click it to jump to that app’s tab and see what the agent did there in context. If the agent switches tabs on its own to act in an app, the active sidebar item updates to reflect the app it is working in.

​

Reviewing and Approving Actions

By default, the agent runs in Normal mode: it works through a task on its own and pauses only before high-stakes actions. High-stakes actions are those that are irreversible or outward-facing: sending a message or email, deleting data, making a payment, or sharing something externally. Routine reads and low-risk writes (creating a draft, updating a record, adding a calendar event) run without a pause. When the agent reaches a high-stakes action, it stops and shows you an Approve action card in the panel. The card has an orange border and contains:

• A plain-language summary of exactly what the agent is about to do.
• Approve (orange button): let the agent proceed.
• Decline (stone button): cancel the action. The agent stops without making the change.
• “Approve and stop asking for this run”: approve this action and let the agent proceed through the rest of the current task without pausing for each step.

Read the card before you approve. The summary is the agent’s description of the action, not a confirmation screen you should click through quickly. If something looks wrong, decline and redirect the agent.

​

Change the Approval Mode

You can change when the agent pauses using the mode dropdown in the composer footer (above the text input):

• Normal (default): pauses only before sends, deletes, payments, and external shares.
• Careful: pauses before every write. Use this when you want to review each step.
• Auto: runs the entire task without pausing. Use this only when you trust the full scope of the task.

Changes take effect immediately, even while a task is running. To understand the full approval model and how the agent classifies high-stakes actions, see How Lava Desktop Works.

​

Stay in Control

You can stop or redirect the agent at any time.

• Decline an action. Click Decline on an Approve action card. The agent stops the current step and returns to the panel.
• Send a new message. Type a correction or redirection in the input bar while the agent is running. The agent reads your message and adjusts.
• Stop the task. Click the stop button in the composer footer to end the current run. Hiding the panel (press Cmd+\ or Ctrl+\) does not stop the task; the agent continues in the background and waits if it reaches an approval step.

In Normal mode, any pending high-stakes action waits for your approval before proceeding. If you are not looking at the panel, the agent holds there until you respond.

​

When the Agent Asks You First

The agent acts on what it can see in your connected apps. When it encounters a reference it cannot resolve from context (an ambiguous name, a missing piece of information, a task that would require guessing) it asks you rather than proceeding on a guess. For example, if you ask it to “email Alex about the proposal” and there are multiple people named Alex in your contacts, it asks which one you mean before composing anything. This keeps the agent from acting on wrong assumptions. If you see an unexpected question, it means the agent needs one more piece of information to proceed correctly.

​

Troubleshooting

If a run fails, stops unexpectedly, or the agent seems stuck, see Troubleshooting.

​

Next Steps