Browser runs: work on sites that have no API
Vocion 5.53 lets an agent do a task in a web browser on a site with no API or connector, while you watch beside the conversation. Anything that would change something waits for your approval.
Status: shipped. #1365, cut as v5.53.0. One migration (0210_browser_runs). Off until you set VOCION_BROWSER_RUNS=on.
A lot of real work sits behind a sign-in with no API: a supplier's portal, a state filing site, a carrier's tracking page. Until now an agent could tell you to go look. Now it can go and look itself, and you can watch it do it.
What you can do now
- Hand a task to a browser from chat. Say "use the browser to find what we owe Fernhill Supply", or type
@browserin the composer. The run opens in the pane beside the chat. - Watch it work. The pane shows the browser live, with each step and the reason for it.
- Approve before it changes anything. A click that would save, send, pay or delete stops the run, with the button outlined on the page, until you press Approve.
- Get the answer as a table. What the run found comes back under the answer, in the order the site showed it.
- Keep passwords out of the model. Mark an input secret and the agent only sees a placeholder. The browser types the value.
- Let it survive a deploy. The run saves a checkpoint after every step. If the app restarts, the run picks up where it was, still waiting for you if it was.
On a deployed installation the browser runs in its own container, apart from the app, and can only open public addresses.
Where to find it
- Ask in any chat, or open Browser runs at
/dashboard/browser. Run the practice task there tries it on a practice supplier portal that reaches nothing real. - Docs: Browser runs.
Get it
Update to v5.53.0 or later and set VOCION_BROWSER_RUNS=on. The AWS compose file adds the browser container and points the app at it. To give browser runs to one agent only, add harness.grantTools: [browser_tasks] to its YAML instead.