Our supervised browser was blocked by a human check
We made our supervised browser the default and reverted it the same day, because a bot-protection provider flags its data-centre addresses and both the agent and the human operator ended up in a verification loop.
Development track
A supervised browser is a real browser session that an agent drives while a human watches the same session and can take control of it. Ours is part of Rove, our browser tool built for exactly that arrangement: the agent does the work, the person can see everything it sees, and the person can step in when a step needs a human. On October 11, 2026, we made that browser the default for supervised sessions. We reverted the change the same day. Inside the session, everything worked. Reaching the site we tested it on was the part that failed, for the agent and for the human in the same session alike.
What happened
Capability grounds were the whole basis of the adoption decision. For work a person oversees, a supervised session is the better tool: one shared view, a clean handoff of control, a record of the session. We set it as the default and ran a comparison against a real site, indiehackers.com, a public site fronted by Cloudflare, a major bot-protection provider.
From the first page load, the two browsers behaved differently. Our previous default, the built-in browser, loaded the site cleanly. Our supervised browser received a Cloudflare 403 page, the "Just a moment…" holding page the provider shows while it decides whether to trust a visitor. Next, the agent met the provider's human check, Turnstile, and looped on it: solve the check, get challenged again. When the human operator took control of the same session through the viewer, the handoff the product exists to provide, the loop was the same. Who was driving did not cause the failure. Anyone arriving from that address was failing it.
The comparison, as a table:
| Browser | Result on indiehackers.com, October 11, 2026 |
|---|---|
| Built-in browser, ordinary egress | Site loaded cleanly |
| Supervised browser, fleet data-centre egress | 403 holding page, then a Turnstile verification loop for the agent and for the human operator |
Same day, the default went back. Supervised sessions returned to the built-in browser, and our supervised browser stayed available for sites that do not challenge it.
The timeline
The timeline, as a table:
| When | What happened |
|---|---|
| October 11, 2026 | Supervised browser is made the default for supervised sessions |
| October 11, 2026, later the same day | Comparison test: built-in browser loads the test site cleanly; supervised browser gets a 403 holding page and a verification loop for the agent and for the human operator. Default is reverted |
| October 13, 2026 | Follow-up issue is opened proposing that browser egress only be routed through a consumer-grade tunnel in proxy mode, with a stated acceptance test |
What was actually wrong
What failed sat outside the browser. A bot-protection provider scores the address traffic arrives from, and data-centre addresses start with a poor score. Our supervised sessions run from a fleet, so their traffic leaves from data-centre addresses, and the provider treated that traffic as machine traffic before any page loaded. Whether an agent was driving, or a person who had just taken control, made no difference to the outcome. Address decided it.
Our adoption decision had a gap of the same kind. We had evaluated the session: how it renders, how control passes to the human, how the work is recorded. Reachability against a protected site had never been evaluated as a condition for making it the default. A default is a claim about every session that follows it, including sessions aimed at sites that challenge unfamiliar addresses. We tested the tool at the point where it is strongest and adopted it for the point where it had never been tried.
One limit on the evidence belongs in the open. We have one observed comparison, on one protected site, on one day. No count exists of how many sites challenge this egress, and this post does not estimate one. A single comparison was enough to reverse a default, because the failure mode blocks the session completely when it occurs. Describing how common the failure is across the web would take more than that one comparison, and we do not use it that way.
What we changed
Back the default went on the same day, and it stays reverted at the time of writing. On protected sites, supervised sessions run in the built-in browser. Where a site accepts its traffic, the supervised browser is used. We describe that arrangement as a workaround, because that is what it is: the built-in browser does the reaching, and sessions that need to be on a protected site give up the supervised session's shared view and control handoff.
Our fix is scoped, as a proposal with an acceptance test, and it is aimed at the egress rather than the browser. Route the browser's traffic only, through a consumer-grade tunnel running in proxy mode, so the session leaves from an address the provider scores as ordinary visitor traffic. All other traffic from the same machine keeps its existing route. Nothing else about the session changes.
Its acceptance test is stated in advance. A supervised session loads the test site with no verification loop, once with the agent driving and once with the human driving, and other traffic from the same machine is unaffected. Until that test passes, the supervised browser does not become the default again. Because the proposal has not shipped at the time of writing, this post reports a reversal and a scoped fix. Resolution is not what it reports.
What it cost
Direct cost was small and we can state it exactly: one adoption, one comparison test, one reversal, all on October 11, 2026. A same-day reversal is cheap as failures go, and catching the problem in our own testing, on the day the default changed, is the outcome the testing exists to produce.
Until the fix ships, cost is functional. Every supervised session that has to reach a protected site runs in the built-in browser instead, without the supervised session's shared view and handoff. Providing that feature set is the reason the product exists, and it stays unavailable on a class of sites we cannot enumerate from one observation. Cost continues until the egress fix passes its acceptance test. No sourced figure exists for hours lost to the loop or to the workaround, and this post does not invent one.
What we take from it
Capability and reachability are separate properties, and a browser needs both before it can be a default. Inside our supervised browser, the session did everything it was built to do. None of that was reachable on a protected site, because the decision that blocked it was taken from the traffic's address before the session began. Any tool an agent uses to act on the web inherits the reputation of the route its traffic takes, and that route is part of the product, whether we designed it as one or not.
For another team, the transferable version is a test, run before adoption rather than after. If a tool will be the default for reaching other people's sites, try it against a protected site first, with the agent driving and with a human driving, and treat a block at that stage as a finding about the tool, not as an edge case to route around later. When the test fails, ship the reversal the same day. A default that strands both the agent and the person supervising it teaches users to distrust the supervision itself, and that trust is slower to rebuild than any single feature on our list.