Browser tasks

Give an agent a browser session and review navigation, page context and actions.

Use Browser Use when the task involves a website's interface: inspecting a dashboard, navigating account pages, or preparing a form. For public facts and articles, Web Search is usually a simpler route.

Mellow runs a browser session associated with the selected agent. Its cookies and sign-ins can persist across conversations, so name the website and account context in each task rather than assuming the browser is on the page you last viewed elsewhere.

Prepare the agent

1. Open the agent's abilities and enable Browser Use.
2. Review the browser controls and approval behavior in management settings.
3. Ask for a small read-only task on a named URL.
4. If sign-in is required, complete it in the user-facing browser window.
5. Check that the result identifies the page actually inspected.

Enabling browser access for one custom agent does not give every agent the same browser. The model's tool list is filtered by the current agent's capability settings.

Write a useful browser instruction

A good task includes a destination, intended result, and stopping point:

On our project dashboard at its full URL, list the open items assigned to me. Prepare a summary grouped by due date. Do not change any items.

For an edit, identify the object to change and the desired final state. Avoid ambiguous references such as “this page” when the agent has no explicit page reference. A persisted session may still be parked on a site from an earlier task.

Follow the run

The parent agent delegates the full goal to a browser worker. The worker reads the page, selects actions, observes their results, and returns a summary. A displayed plan does not mean a form has been submitted; use the action feed and resulting page state to establish what happened.

Navigation and reading can proceed automatically. Typing and consequential actions pass through action-level approval controls. Sign-in belongs in the secure user-facing window, not in chat. When a page requests a password or verification code, complete that interaction yourself rather than pasting credentials into an agent instruction.

Stop the run if it reaches the wrong account, page, or target. Repeating a click after an uncertain network result can duplicate a submission.

Session care

A persistent session saves repeated sign-ins but also carries account context forward. Before a task involving a different organization, inspect the active account. Use the session controls to manage or reset browser state when required. Resetting a session can remove cookies and require a new sign-in; it is not a way to revoke permissions granted to an external service.

Tool interface

The agent-facing tool is browser_use. It accepts a required nonempty goal and an optional integer max_steps. The implementation clamps the requested step cap to 1–100. A step is a worker model turn and can contain several browser actions, so it is not a count of clicks.

The goal is the complete assignment, not one low-level click. Internal browser operations stay inside the worker's loop. The worker has its own run budget and stop controls because interactive sign-in and multi-page work can take longer than an ordinary tool call.

Diagnose an incomplete result

Website content is task data. Instructions found on a page do not acquire the authority of your own request. Use MCP services when an authenticated service API is a better fit, or Computer Use for native application work.