Anthropic has built its own browser into Claude Cowork, which runs inside the desktop application.If a task requires a website, a browser area opens in a side window in which the agent visits pages, reads their content, clicks and enters text.
In practice, this means: filling in forms, reading values from a dashboard, navigating through multi-step interfaces.And it does so even when the target system does not offer a programming interface — that is, no technical access option through which other software could directly retrieve or write data.
Why this is interesting for legacy systems
In many mid-sized companies, automation fails not for lack of will but for lack of interfaces.Supplier portals, government interfaces, booking platforms, older industry solutions in the browser: anyone who needs data here works with copy-and-paste or with screenshots and subsequent re-entry.
Until now there were two ways of doing this.Either classic robotic process automation, i.e. scripts that repeat recorded click paths — robust as long as nothing changes in the layout, and maintenance-intensive as soon as something does.Or an integration project that first creates an interface, which in the case of a third-party portal is usually not within your own control at all.
An agent that reads the interface and derives from it where it needs to click shifts this boundary.It does not need an exactly predefined path, but a goal.This makes it more tolerant of small changes — and at the same time less predictable than a script.
What this demands in terms of care
An agent that types and clicks in real interfaces acts on behalf of the logged-in user.Anyone who wants to use such processes in production should clarify a few points in advance:
- Credentials and permissions: Which account does the agent work with, and what is this account allowed to do?A dedicated access with minimal rights is better than the admin login of the department.
- Writing actions: Reading is harmless, submitting is not.For bookings, orders or status changes, confirmation by humans or a clearly limited task definition is needed.
- Traceability: What did the agent do and when?Without a log, troubleshooting after a week is hardly possible.
- Data protection and terms of use: Personal data that passes through an agent requires the same consideration as any other processing.And some portals prohibit automated access.
- Reliability: An agent that is right in nine out of ten cases is usable for research and not for invoice approval.
The other direction: building your own applications agent-friendly
More exciting than the question of what agents do with third-party portals is, for many projects, the reverse: How do you build a web application that can be well served by humans and by software?Because if agents use interfaces more frequently in future, your own application becomes the counterpart.
Much of this is not a new discipline but solid craftsmanship that now gains a second benefit:
- Semantic HTML and clean forms: Labelled fields, meaningful field types, real buttons instead of clicked divs.What a screen reader understands, an agent understands too.
- Stable structure: If labels and orders shift with every release, any automation breaks — both the machine one and the one in users' heads.
- Predictable states: Clear feedback on success and error, unambiguous loading states, no silent redirects.
- Fewer tricks: Endlessly reloading lists, elaborate drag-and-drop interactions and purely visual cues are hard for agents to grasp.
- An interface anyway: Where there is an API, it is the better way.An agent that operates interfaces is the way out for everything that has none — not a replacement for proper integration.
For TYPO3 projects this means concretely: editorial processes, approvals and imports can in part be supported by agents if the backend processes are clearly structured.And in the frontend, what already applies still applies — accessible, semantically correct output is the best starting point for everything that follows.
Realistic entry scenarios
Anyone who wants to try out such automation should start small and begin where a mistake costs little.Tasks that arise manually and regularly today but are not time-critical are well suited: pulling together key figures from several portals into a weekly overview, cross-checking data sets between two systems, bundling status queries.
Less suitable is anything that directly moves money or is legally binding, as long as the processes are not safeguarded and logged.A sensible intermediate model: the agent researches and suggests, the human confirms.This significantly reduces the effort without responsibility slipping.
Assessment
An agent with its own browser lowers the barrier to entry for process automation where interfaces are missing — and in mid-sized companies that is the norm, not the exception.For web and software projects, this shifts a quality characteristic: clean semantics, stable structures and documented interfaces are no longer just a question of accessibility and maintainability, but also help determine whether an application can be integrated into automated processes at all.
Anyone building new today should consider both user groups — people and software.
Sources
- Anthropic is building its own browser into Claude Cowork — The Decoder