A browser built for people is a rough workplace for agents

Most web agents still do a strange imitation of office work. They inspect a page, click something, look again, type, scroll and repeat.

That is understandable. The web was built for people, and a browser is the common doorway. But a long chain of tiny actions is slow, expensive and easy to knock off course.

A position paper presented at ICML 2026 argues that stronger models will not solve this on their own. Researchers from Microsoft, Ohio State University and the University of British Columbia say the web also needs a better action layer.

Their proposal is called Web Verbs. Instead of asking an agent to discover the sequence of clicks for searching a hotel every time, a site or developer could expose one documented action with clear inputs and a structured result.

The idea is simple: let the agent reason about the job, not the location of the next button.

A verb is more than a renamed API

A web verb looks like a typed function. It has a name, defined inputs, a predictable output and a description of what it is supposed to do.

It could call an official server API. It could also wrap a carefully maintained Playwright workflow when the action only exists in the browser. The agent sees the same high-level contract either way.

The paper gives hotel search and map directions as examples. A directions verb can accept two places and return distance, time and route as fields. A later step can use those values directly instead of reading them from a page and hoping the meaning survives in the conversation history.

The contract can carry preconditions, postconditions, permission tags and logging hooks. That creates obvious places to check whether a user is signed in, whether an action needs confirmation and whether the result matches the request.

This does not remove browser automation. It moves the fragile part into a reusable component that can be tested and repaired once, rather than rediscovered by every agent on every run.

Two examples show what gets lost in the clicks

The authors compare verb-based programs with a browser agent on two tasks.

In the first, the user asks for hotels ranked by the total distance to a set of museums. The verb-based agent writes a short loop, asks for each distance and adds the results. The browser agent instead builds one route through the museums. The answer looks plausible, but it solves a different problem.

In the second, the user wants one item from each furniture category while keeping the whole list under $1,000 and maximising ratings. The verb-based program evaluates complete combinations. The browser agent makes local choices, crosses the budget and leaves categories empty.

The team says its wider prototype implements verbs for more than ten websites and completed over 100 complex tasks. The code and verb catalogue are public.

Those results come from the researchers’ own prototype. The paper is a position paper with representative case studies, not a large independent comparison across current commercial agents.

The difficult part moves to the ecosystem

Typed actions make an individual workflow easier to inspect. They also create a new maintenance and governance job.

Someone has to choose the right size for each verb. A single checkout action may hide choices a user should see. A hundred tiny shopping actions bring back the same long trace the proposal is trying to avoid.

Sites need a stable way to publish and name verbs. Agents need a reliable way to find the right one. Permission rules must cover authentication, personal data, consent and actions with financial or legal consequences.

The authors suggest a domain-controlled naming system, stable public element locators and safety metadata. They also allow for human confirmation before consequential steps. None of that is a web standard today.

There is an economic question too. Sites may welcome reliable automation, or worry that agents will bypass advertising, design choices and customer relationships. A technical interface does not settle that argument.

What is demonstrated, what is proposed and what remains open

The demonstrated work is a public prototype, a catalogue of typed actions and a set of tasks. The paper shows that an agent can compose those actions into short programs with visible data flow and constraints.

The broader claim is that this pattern will be more reliable, efficient and auditable than letting agents drive every site one click at a time. The case studies support the direction, but they do not establish how large the gain would be across the open web.

Web Verbs are not a standard and major websites have not agreed to publish them. Independent benchmarks, security models, versioning rules and incentives for long-term maintenance are still missing.

The proposal is useful because it changes the question. Instead of asking only how to make an agent better at seeing a page, it asks what kind of web an agent should be allowed to act on.

For many routine jobs, fewer clicks may be the point. The harder work is making the larger actions trustworthy.

Sources

  1. ICML 2026 paper — Web Agents Should Use Typed ActionsPrimary position paper. Source for the Web Verbs design, prototype scope, case studies, reported results, limitations and open problems.
  2. Microsoft Research — Publication pagePrimary publication record confirming the authors, ICML 2026 venue and the proposal’s stated scope.
  3. Microsoft Research — MSR-Web-Verbs repositoryPrimary public prototype, verb catalogue, task materials and implementation documentation.
  4. ICML 2026 — Conference downloadsPrimary conference index listing the paper in the 2026 programme.
  5. Microsoft — Introducing NLWebPrimary background on NLWeb, the content-oriented semantic layer that the Web Verbs proposal extends toward actions.