Navigation agent

The navigation agent is Trace's browser automation system. Describe a task in plain language in the chat sidebar, and it plans and executes the steps directly inside the live browser view, adjusting on the fly if a page doesn't behave as expected.

How it works

1. Plan: given your instruction, the current URL, and the visible page text, the agent generates a step-by-step plan.
2. Execute: each step runs directly in the browser. Cookie consent banners are dismissed automatically before any interaction.
3. Observe: after every step, the agent snapshots the page (URL, title, text) and decides whether to continue, replan, finish, or fail.
4. Adapt: if the page changed in a way the plan didn't expect, the agent replans the remaining steps instead of getting stuck.

What it can do

The agent can open pages, click elements, type into fields, scroll, wait for content to appear, choose dropdown options, check boxes, hover, send keyboard shortcuts, and highlight file inputs for you to complete manually (file selection itself always requires you). It can also search your indexed browsing history, retrieve a stored page's full content, and snapshot data across multiple pages for later comparison.

What it's good at

Comparing prices or information across sites. Ask something like *"compare the price of a specific product on three retailers"* and the agent visits each site, extracts the relevant data, and returns a comparison once it's checked all of them.

Filling out and submitting forms. A single instruction like *"send a contact form to a company's site with my name, email, and a message"* fills every matching field (by label, placeholder, or name) and submits it.

Collecting paginated data. The agent can page through search results or listings, accumulating data from each page until it hits a limit or runs out of pages.

Updating account settings. Multi-field settings forms (a display name and bio, notification preferences, a timezone) get filled and saved in one instruction.

Credentials

If a task requires logging in, the agent will ask you directly for a username and password rather than guessing or storing anything in the plan itself. Credentials you provide are used in memory for that step only and are never logged. You can also store credentials ahead of time in the credentials vault for the agent to use automatically.

Limits

• A task is capped at 18 observation cycles, so a page that keeps changing unexpectedly can't loop forever.
• Page text sent to the agent is capped in size (roughly 1,500 characters per observation, 10,000 for planning) to keep it fast and focused.
• Memory-only steps (searching your history, retrieving a stored page) skip the browser entirely and don't count against observation limits the same way.

Warning:

The navigation agent acts on real pages, including submitting real forms. Review what it's about to do before approving anything that isn't purely read-only.

Next steps

• AI sidebar & skills for the skill chips that trigger common navigation-agent tasks.
• Credentials vault to store login details ahead of time.