Websites are long since no longer accessed only by humans and classic search engine crawlers. Increasingly, AI agents are being added: programs that, on behalf of a user, independently access pages, gather information, fill out forms or initiate processes. The strategy team of the W3C, the standards body for web technologies, has outlined requirements for future web standards in this regard.
The core idea can be summarised briefly: a mandate to an agent is not a blank cheque. Anyone who sends out an agent delegates a task – not the entire decision-making authority over data, accounts and transactions. For standardisation, this means that control must be technically traceable and representable and must not lie solely at the discretion of the respective provider.
Why the topic is now on the agenda
The web has historically grown around two forms of use: humans in front of a browser and machine access via interfaces. Agents lie in between. They use the interface built for humans but act machine-like, at high frequency and often without anyone tracking every single step.
This raises questions that existing standards only partially answer: how does an operator recognise that an access originates from an agent? On whose behalf is this agent acting? And how far does this mandate extend? As long as these questions remain open, operators resort to improvised means – from blocklists to bot detection, which can also slow down legitimate use.
Four levels on which operators need answers
Regardless of which concrete specifications ultimately emerge, the problem areas can be sorted well. They affect every project that processes transactions or personal data:
- Identity: Is this a human, an agent or an agent acting on behalf of a human? Without a reliable distinction, neither rights nor liability can be clearly assigned.
- Scope of mandate: An agent meant to compare prices does not need authorisation to purchase. A distinction must be possible between reading, writing and binding completion.
- Traceability: Which actions has an agent performed, at what time, on what basis? Without logging, nothing can be reconstructed in the event of a dispute.
- Revocation: Granted permissions must be revocable at any time, and effectively for ongoing processes, not only for future ones.
What this means for the design of websites
Part of the work is standardisation and will take time. Another part is already in the hands of operators today. Those who structure their applications cleanly make them usable for agents without giving up control.
Above all, this includes separating information from transaction. Content that is public anyway – service descriptions, opening hours, technical data – may be well accessible to machines. Processes, on the other hand, that cost money, trigger contracts or process personal data should be tied to explicit authorisation and not merely to someone being logged in somehow.
Equally practical: semantically clean markup. A page whose structure is understandable even without visual interpretation can be operated more reliably by agents – and is at the same time more accessible. The same care that supports assistive technologies pays off twice here.
Caution against premature commitments
What the future standards will look like in detail is open. Accordingly, it is advisable not to commit architectures today to a particular protocol or a particular agent ecosystem. More sensible is decoupling: business logic and permission checks belong in a layer that works independently of whether the access comes from a browser, an app or an agent.
Provider-specific integrations can then be swapped without the core application having to be rebuilt. This is not a new insight, but it gains weight when an additional class of access is added whose rules of the game are still being negotiated.
Do not delegate organisational questions to technology
Not everything that needs clarifying here is a technical problem. Who in the company may determine which agents may trigger which processes? How is it documented that consent was given? Who regularly checks whether granted permissions are still needed?
These questions resemble those that have long been asked about interfaces and service accounts. The difference lies in the scope: an agent decides on the intermediate steps itself within its mandate. All the more important is that the mandate itself is precisely outlined.
Assessment
For web and software projects, the view of the target group thus shifts: alongside humans, agents are a form of use that should be considered in architecture, permissions and logging. Anyone who today implements clean semantics, clearly delimited permissions and traceable logging is well prepared for coming standards – and incidentally improves maintainability and accessibility. Waiting for concrete specifications is legitimate; leaving one's own transaction paths uncontrolled is not.
Sources
- W3C: Ein Auftrag ist kein Freibrief für KI-Agenten — heise developer News