Until now, the roles were clearly divided: a human opens a website, clicks through navigation and forms, logs in, downloads something.AI assistants at most summarised texts or reproduced content they had previously read in.With the built-in browsers of Claude Cowork and ChatGPT Work, this boundary is shifting.Both systems work in an agentic process – that is, they execute several steps independently in sequence – directly on websites.Switching to a separate browser window is no longer necessary, and the assistants also handle logins.
Cowork additionally comes with memories that are linked to the respective chat.The assistant therefore keeps track of what was previously discussed and done over the course of a work history.
For users, this is primarily convenient.For everyone who operates websites, it is something different: a new user group that behaves differently from humans – and that you can neither see nor welcome.
Why this is more than just another crawler
Search engine crawlers read pages.An agentic browser operates them.It clicks, fills out, logs in, follows processes across multiple steps.This means it no longer affects only the content level of a website, but the interaction level: forms, booking flows, customer portals, download areas.
This has a practical consequence.A page can be excellently prepared in terms of content and still remain unusable for an agent because the decisive step cannot be reliably executed.Conversely, the same applies: what an agent can operate cleanly is, as a rule, also well usable for people with assistive tools.
Structure and semantics become the interface
An agent has no eye for layout.It cannot infer from size, colour and position that a particular element is the submit button.It depends on what is in the markup – that is, in the HTML code of the page.
As a result, fundamentals that many projects have treated rather casually over the years gain weight:
- Semantic elements instead of generic containers: A real button, a real form field, a real heading hierarchy tell a machine what something is.A div made interactive with scripts says nothing at all.
- Labelled form fields: A field whose meaning emerges only from a text graphically placed next to it is a guessing game.A correct association of label and input field is not optional.
- Traceable states: Error messages, loading states and confirmations must be recognisable in the code, not merely visible as a red border.
- Stable flows: Processes that depend on exact mouse positioning, hover effects or drag-and-drop are difficult for agents – and, incidentally, for many people too.
In short: accessibility and agent-friendliness largely coincide technically.Anyone who takes the requirements of accessibility seriously has already done a large part of the preparatory work.
The delicate part: login and forms
The fact that the assistants also log in is the point at which things become organisationally demanding.A login means access to personalised areas, to order data, contracts, documents.When an agent acts on behalf of a person, from the system's point of view it is initially indistinguishable from that person.
Operators should therefore ask themselves a few questions early on, rather than having to answer them in the event of a disruption:
- Which areas of your own portal should be operable automatically at all – and which deliberately not?
- How do existing protective mechanisms such as bot detection, rate limits or captchas behave when a legitimate agent works on behalf of a real customer?
- Where do actions with legal or financial effect arise that require deliberate confirmation by a human?
- How are accesses logged so that it remains traceable afterwards what happened?
The answers differ depending on the industry.A product catalogue may gladly be searched by machine.A cancellation flow or a payment approval requires a different level of care.
What this means for operations
Alongside security and semantics, there is a third level that is easily overlooked: analytics.If a growing share of interactions runs via agents, metrics change.Time on site, click paths and bounce rates then partly measure the behaviour of software rather than the behaviour of a human.Anyone basing decisions on such figures should at least know that the composition of traffic can change.
The preparation of content also deserves a second look.Information that today is contained only in images, PDFs without a text layer or in elaborately animated sections is difficult for an agent to access.Clear texts, clean data structures and machine-readable markup increase the chance that your own content is reused correctly – instead of approximately or not at all.
No reason to panic, but reason to take stock
How quickly agentic browsers will become established in everyday use cannot currently be quantified seriously.The only certainty is that the function is no longer experimental and on the margins, but is being built into the working environments of major providers.That is enough of a reason to examine your own site in this respect – without rebuilding everything.
A pragmatic start: go through the two or three most important user paths on your own website and check whether they work without a mouse, without knowledge of the layout and without interpretation of images.
What stands out there is usually also what makes agents fail.
For web and software projects, this above all means: investing in clean, semantic markup and in accessible operating flows now pays off twice.It fulfils regulatory requirements for accessibility and at the same time determines whether automated assistants can work with an offering.Anyone planning relaunches or new portals should incorporate these requirements into conception and acceptance right from the start, instead of retrofitting them later.