Agents that click: n8n Browser Use, Dust's extensions, Grok Bot's Vault, computer-use models, and the governance gap
n8n opened Browser Use to everyone, Dust gave its extensions page awareness, Grok Bot added a Vault, and Haiku 5.5 made computer use cheap. The browser agent is now a feature of tools you already own. Here is the governance that has to catch up.
In one month, four of the tools we cover most gave their agents a browser or made the browser they already had more capable. n8n opened Browser Use to every user. Dust put page awareness into its Chrome and Firefox extensions. Grok Bot added a Vault so a Bot can sign in and check out on its own. Anthropic's Claude Haiku 5.5 (7 October) brought computer use to a model priced at $0.10 per million input tokens, which makes "an agent clicking through a portal" cheap enough to run at volume.
We wrote in August about browser agents versus workflow automation as a what-to-buy-first question. This is the follow-up: the browser agents have arrived inside the tools you already own, and the governance has not.
What each vendor shipped
| Tool | What it can now do in a browser | Where control lives |
|---|---|---|
| n8n 2.43+ | A Browser Use step inside a workflow: open, navigate, fill, extract | Admin node rules (2.44) can restrict who may add it; personal credentials in team projects (2.46) |
| Dust | Agents aware of the page you are on in Chrome and Firefox; write back to Drive | Data-access permissions per space; publish approval |
| Grok Bot | A persistent cloud computer with browser, plus Vault-stored logins and cards, email in, shopping cards, fewer approval prompts | Enterprise tier: access, network and audit controls; per-user isolation. Bots on one account share a computer |
| Claude Haiku 5.5 and similar | Computer use at commodity prices for your own agents | Whatever you build |
Why this is the trending topic
Browser agents solve the problem SMEs actually have: the work is in portals with no API. Supplier sites, HMRC and council portals, insurer and lender dashboards, the legacy ERP front end. Until now the honest advice was "wait for an API or hire a person". Now the advice is "an agent can do it", and the next sentence has to be about what it is allowed to do.
The governance gap is specific:
- Credentials. A browser agent needs a login. Shared logins with no owner, pasted into a Vault or a workflow credential, are the new shadow IT
- Untrusted content. A page, an email or a PDF the agent reads can contain instructions. Prompt injection via a web form is not theoretical for an agent that can click "Submit"
- Spend. Grok Bot can hold a card. n8n can fill a checkout. Where is the limit, and who set it
- Evidence. When a browser agent does something wrong, the trail is screenshots and a transcript, not a database row. Keep them
- Identity. Systems see a human user logging in, not an agent. See your agents need logins, not API keys
A UK operator's checklist for browser agents
- Allow-list the sites. A browser agent for supplier portals does not need the open web. n8n node rules, Grok Bot Enterprise network controls, or a proxy
- One login per agent, with an owner. Never the finance lead's own credentials. Where the portal allows a dedicated user, create one and log its actions
- Read-only first. Extract, summarise and report for a fortnight before you let the agent submit anything
- Approval on the verbs that cost money. Submit, pay, send, delete, unsubscribe. Human by default; promote to automatic after ten clean supervised runs
- Virtual cards with limits for any agent that can check out. Prepaid where possible
- Treat inbound content as hostile. Emails to a Bot's address, uploaded PDFs, scraped pages. Summarise before acting; never act on an instruction found inside content
- Screenshot and transcript retention aligned to your data retention policy, stored where the agent cannot edit them
- Kill switch drill. Revoke the agent's sessions and rotate its credentials in under five minutes, and name who does it
- Review the DPIA. A browser agent touching customer records in a portal is a new processing activity
Which tool for which job
| Need | Prefer |
|---|---|
| Portal task inside a larger deterministic process (pull a statement nightly, reconcile, alert) | n8n Browser Use behind IF checks and approval nodes |
| Knowledge work on the page you are reading, grounded in company docs | Dust extension |
| Open-ended multi-portal work with lasting context, like a person at a desk | Grok Bot, Enterprise tier, Vault scoped tightly |
| Your own product feature | A computer-use model such as Haiku 5.5 behind your own guardrails |
Bottom line
The browser agent stopped being a separate product category in October 2026 and became a feature of the tools you already run. That is the useful outcome and the risky one. The vendors have shipped some controls (n8n node rules, Grok Bot Enterprise, Dust permissions); the rest is on you. Allow-list, give each agent its own login and owner, keep approval on anything that spends or sends, and keep the evidence. Then let it click.
Related reading
- AI browser agents vs workflow automation: what to buy first
- n8n in October 2026: Browser Use, node rules, durable scheduler
- Grok Bot in October 2026: Pages, Vault, Bot email
- Dust in autumn 2026
- Your agents need logins, not API keys
Sources: n8n Browser Use integration · n8n release notes · Dust changelog · Grok Bot changelog (community-maintained) · Claude Haiku 5.5 pricing coverage · Okta Agent SSO
Erhan Timur, Founder, Digital by Default
Enjoyed this article?
Subscribe to our Weekly AI Digest for more insights, trending tools, and expert picks delivered to your inbox.