The Browser Is an Agent Surface Now

Most enterprise work still happens in a browser. The CRM, the ERP, the HR system, the invoice queue: all of them are web applications behind a login form. An agent that cannot use a browser cannot do that work. An agent that can use a browser, on a consumer setup with saved passwords and a person’s full session, can do far more than that work.
Island made that problem a product category on October 29. In Island’s announcement of its Enterprise Browser for Agents, the company describes a browser built for agents to access, read, and act on SaaS applications, internal applications, and public websites under the same enterprise controls and policies the company already applies to people. Sales agents update CRM records after calls. HR agents help onboard new hires. Finance agents process invoices. Island’s stated examples are ordinary back-office work, which is the point. The browser is not a side channel for agents. For a large share of enterprise tasks, it is the workplace.
One sentence in the announcement carries the design. Island says it governs browser activities, not agents. That is the right instinct, and it matches a rule I keep coming back to in this series: control the point where an action takes effect, not the prompt that produced the action. A policy written into an agent’s instructions can be talked around, misread, or lost in a long task. A policy enforced by the session the agent is working in either allows the click, the form submission, and the download, or it does not. I made the enforcement version of this argument in An Alliance Is Not a Control Plane: count controls where an agent can actually be blocked. A governed browser session is one of those places.
The announcement lists controls that map to problems I have already covered from other angles, which tells me the category is converging on the same short list.
Access is granted per agent and per application, down to which page operations are allowed and whether the agent can read or also write. Read-only mode on sensitive systems is a named feature: an agent can view and extract from an ERP or another system of record without modifying, submitting, or deleting data. That is the permission review from The Permission Is the Policy applied at the level of a single page action instead of an account. A domain allowlist says an agent may visit the finance system. An operation grant says it may read an invoice and may not approve one. The second statement is the one an auditor can test.
Credentials stay out of the model. Island’s built-in vault injects credentials directly into login forms, including applications with no identity provider integration, so passwords never pass through the model’s context window. Read that against The Log Is Also the Breach. Anything that enters a model’s context can end up in a trace, a transcript, or a vendor log. A password that never enters the context cannot leak from it, and a credential the agent never sees cannot be handed to a convincing web page that asks for it.
Attribution is split between person and agent on every event, and sessions can be recorded in full for playback of every page, click, and form submission. That answers a question that has run through several posts in this series, including There Is No Device to Enroll: when work arrives under a person’s name, can you later show which actions the person took and which actions an agent took in that person’s session? If the record cannot make that split, your incident review starts with a guess.
The weak point in the design is also stated plainly, and you should not skim past it. Agents inherit the user’s enterprise identity and operate under the user’s permissions. Island adds that policy can apply to the user alone, to the user and their agents together, or to the agents alone. That third option is the whole ballgame. Inheritance as a default is how an agent ends up with a person’s full reach on its first day, including reach the person rarely uses and never reviewed. I covered the delegation version of this in Delegation Is Not Inheritance: a grant passed down should get narrower, not copy itself. A browser product that lets you set agent-only policy gives you the tool. It does not do the narrowing for you. Somebody still has to write the agent policy, application by application, before the agent starts work.
There is also a cost point buried in the announcement that deserves more attention than it will get. Island argues that screenshot-driven browsing burns tokens, because the model has to interpret pixels and raw page markup at each step, and that its own annotation-based page parsing cuts that consumption. Put aside the vendor claim and keep the arithmetic. Browser work is a chain of small steps, and every step is billed. I made the budget version of this point in Every Agent Needs a Budget: a task with no per-task cap will find its own cost. A browser agent that takes forty steps to file one form has a cost problem and a control problem at the same time, because a long chain of page actions is also a long chain of chances to do the wrong one.
So test any governed browser offer, Island’s or another vendor’s or a setup you build yourself, the same way you would test any other control point. Four tests will do it.
First, the separate policy test. Set an agent-only policy that is narrower than the person’s policy on one application, then run the agent and watch it hit the narrower line. If the tool cannot express a difference between the person and the person’s agent, you have session logging, not agent governance.
Second, the operation test. Grant read access to a system of record and deny the submit action. Then have the agent try to complete a real task that ends in a submission. The denial has to happen in the session, with a record, not as an error message the agent can route around by opening a second tab or a different tool.
Third, the credential test. Pick one application with no single sign-on, where a vault has to inject the password. Search the agent’s transcript, traces, and session record afterward for that password. If it appears anywhere outside the vault, the isolation claim failed in the only place that matters.
Fourth, the boundary test. The browser covers one surface. Agents also act through APIs, MCP tools, desktop applications, and scheduled jobs that never open a page. Ask which of your agent’s tools fall outside the browser session, and write that list down before you count the browser as a control. A governed session next to an ungoverned API key is Egress Is the Boundary restated for a new surface: the work will flow through the opening you did not close.
Island’s CTO, Dan Amiga, is quoted in the announcement predicting that within a few years most enterprise browser activity will come from agents, not people. I cannot verify that forecast, and I do not need it to be exact. The direction is already visible in the work agents are being handed today. If agents are going to work in the browser, the browser session has to carry identity, a permission grant, credential handling, and a record that were designed for an agent, not borrowed from the person who launched it.
Thinx-Tank