Agentic Web & Architecture
Clicking Is Obsolete. Web Agents Need a Different Language
mlj.solutions
mlj.solutions AI agenti dnes klikají po webu jako člověk. A právě proto jsou nespolehliví. Web Verbs navrhují jiný přístup: místo slepého klikání dostane agent typované akce. add_to_cart. book_flight. cancel_subscription. Každá ověřitelná. Každá auditovatelná. Méně kroků. Víc kontroly. Žádné překvapení. 👉 Jak moc vás trápí křehké automatizace ve vašich procesech?
View on Instagrammlj.solutions
mlj.solutions 50 kliknutí. 3 příkazy. Stejný výsledek. ⚡ AI agenti dnes v prohlížeči klikají jako člověk - a právě proto jsou nespolehliví. Jedno přesunuté tlačítko a celý proces padá. Web Verbs navrhují jiný přístup: místo slepého klikání dostane agent typované akce. add_to_cart. book_flight. cancel_subscription. Každá ověřitelná. Každá auditovatelná. Méně kroků. Víc kontroly. Žádné překvapení. 👉 Jak moc vás trápí křehké automatizace? Celý příběh ve čtvrtek. 🎬
View on InstagramWhat the Study Is About
Today's web agents control the browser through clicks, keystrokes, and DOM manipulation. The study argues that this layer is too fragile and unauditable, and proposes Web Verbs — a semantic layer of typed actions that websites expose to agents as a stable interface.
35.4%
top agent's success rate on multi-site tasks
IBM CUGA achieved only 35.4% success on tasks spanning multiple websites, despite exceeding 60% on simple tasks.
100+
complex tasks completed without failure
In a prototype test with Web Verbs implemented, every task was completed successfully.
1
typed function
A task like 'find a hotel and sort it by distance from museums' shrinks to a short program with Web Verbs, with no clicking through the DOM.
What This Means
- Clicks are semantically empty, the agent doesn't know what a click means — only what happens on the page after it.
- Errors accumulate over long sequences, the longer the chain of steps, the higher the chance of failure.
- Web Verbs are auditable, every call has inputs and outputs that can be logged and checked.
- Web Verbs can be shared and reused, one developer writes an implementation that works for every agent.
- Stronger models won't solve this, without a semantic layer, even better models will suffer the same structural problems.
Our Take
A good agent on a bad interface still fails. Just as we adapted the web for mobile, we now need to adapt it for agents.
Michal Dobrovolný
Founder & CTO
Zdroj
Jiang, L., Xi, R., Liu, Z., Chen, S., Lin, Z., & Nath, S. (2026). Web Verbs: Typed Abstractions for Reliable Task Composition on the Agentic Web. arXiv:2602.17245. https://arxiv.org/abs/2602.17245
Keep reading
Don't miss these posts
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
