Agents are learning to use the web, and the web is learning to charge them
On 3 October, developer Marko Denic published a guide to WebMCP, a Chrome experiment that turns an ordinary HTML form into a callable tool for AI agents using three attributes. The post lands in the middle of a wider argument about who controls browser automation.
On 3 October, developer Marko Denic published a walkthrough of WebMCP. It is a Chrome experiment that lets a page expose its features to AI agents as named tools with typed inputs. The agent no longer has to guess its way around a rendered interface. According to Denic, a simple search form needs three attributes: toolname, tooldescription and toolautosubmit.
His framing is blunt. An agent driving a normal site "looks like someone working through a mouse with oven mitts on." It reads the screen and guesses which box is the search field. It guesses which button submits, then reads the screen again to check. Every guess is a chance to be wrong.
WebMCP is still a trial, not a standard
Denic is careful about the limits. WebMCP is an experiment in Chrome, and shipping the attributes is not enough. Without an origin trial token, document.modelContext only exists for users who have flipped a flag. The trial he cites runs on Chrome 149. Browsers that do not know the attributes ignore them, so everyone else gets the page as it always was.
The more interesting detail is where the tool lives. For forms, the browser derives the schema from the markup. For anything else, a calculator or a configurator, the page registers the tool from JavaScript. That takes a name, a description, an input schema and an execute function. Denic's example wires execute to the same function the sliders call, so a human and an agent take one code path. He also recommends leaving toolautosubmit off for anything that sends, buys or deletes, so the user still confirms.
That is one answer to the browser automation problem: make the page legible instead of making the agent smarter. It is not the only one in circulation. It is also not the one with the most money behind it.
The harness is where the work is going
Tim Waldin's post from 4 October argues the binding constraint is not perception. It is where the output of an agent's work ends up. He quotes Andrej Karpathy, who wrote on Friday that "We'll be spending a lot more time trying to understand the outputs of language models." Karpathy added that more work will "rise up the abstractions into oversight and understanding." Waldin's complaint is that a diagram or an HTML explainer from an agent scrolls away in a chat transcript. The representation of the system is then never quite part of the conversation about it. His app, easl, gives each repository an infinite board where terminals, code tiles, browsers and notes sit together. A hyper-key click adds whatever you point at to the next prompt.
That is a design claim, not a benchmark, and it should be read as one. But it matches what practitioners report elsewhere. Alexey Indeev wrote on 4 October that he merged 562 pull requests in two weeks while spending four of those nights hiking. His workflow now turns on a coordinator agent that does no work directly and delegates to sub-agents. Tests, CI and review bots gate the merges. He is explicit that the models changed recently, Opus 5.5 in particular. A couple of weeks earlier, he would not have recommended the approach.
What the state checks say
The most concrete number in this batch comes from Microsoft and Hugging Face. A joint post published on 3 October describes ThinkingBox, a benchmark of 507 stateful business workflows. Each runs 20 times against various models. It grades the terminal backend state and side effects an agent leaves behind, not the sentence it returns. In a common-set ablation covering 121,680 valid trials across 12 models, 79,853 attempts failed the executable checks. Of those failures, 67.24% still terminated cleanly, invoked a state-changing tool and reported no final tool error.
The illustrative case is a $745 kitchen appliance stuck in a courier exception at a Nashville distribution center. It was fifteen days past its estimated delivery date. The agent made nine tool calls and read the refund policy twice. It correctly concluded the customer did not qualify for late-delivery compensation. Then it closed the ticket as resolved when the required end state was on hold, and never answered the question she asked. Executable checks found wrong field values in 77.61% of the clean-looking failures. They found unintended extra effects in 43.30% and missing required effects in 25.36%.
A trajectory is a claim. Database state is the evidence. Repetition is the trust test.
That line, from the Microsoft and Hugging Face post, is the sharpest thing written about agent evaluation this week. It also undercuts a lot of demo culture. A well-formed tool call is not the same as a correct outcome. One correct run is not the same as a reliable one.
Permissions and drift
Apple moved on the permission side. In a blog post on its developer site, reported by 9to5Mac on 2 October, the company said some developers are using Full Disk Access in risky ways. That setting is meant to let backup apps work. Apple said the misuse could expose "everything on their systems, including files, mail, messages, and even browsing history, without users' full knowledge and understanding." Apple said it will add controls so that granting that access requires "very explicit user action." It tied the change to AI agents becoming "increasingly capable and autonomous." It did not say what the controls are or when they ship.
Meanwhile, the failure mode that no permission dialog catches is drift. The sdarchitect blog published the fourth part of a series on AI risk on 3 October. It asks why an agent's behaviour changes when nobody changed the agent. No prompt was deployed and no code was touched. The change-control system is clean. Yet the agent that worked last month now escalates decisions it used to handle. The post attributes this to the whole system of models, harnesses, tools and other agents. It cites NIST guidance on post-deployment monitoring. It argues that if behaviour can change continuously, evaluation has to run continuously too.
A paper on arXiv from June, by Emre Turan, adds a number that should make anyone designing approval gates pause. On a hand-labeled set of 125 adversarially weighted agent actions, reviewers agreed only moderately on what counts as risky. The Fleiss kappa was 0.52. When the reviewer is modelled as fatiguing under escalation load, realized safety becomes an inverted U in the escalation rate. More human oversight can make a system less safe, and the safety-optimal guard escalates below full escalation.
None of this makes browser automation less useful. It does mean the interesting engineering has moved away from making an agent click a button. Denic's three attributes, Waldin's shared board, Indeev's merge queue and the ThinkingBox state checks all try to put the same thing around the model. That is a surface it can act on, plus evidence that the action landed. The pages, and the agents, are still arguing about who provides it.
Sources
14- 01Let AI agents use your site with 3 HTML attributesEN
- 02your agent's diagrams need somewhere to liveEN
- 03I stopped reviewing my agents' code. Here's what I do insteadEN
- 04The Agent Said It Was Done. The Database Disagreed.EN
- 05Apple says it's tightening macOS privacy controls amid the rise of AI agentsEN
- 06AI Risk - A user's guide (Part IV): Understanding Agent DriftEN
- 07Oversight Has a Capacity: Calibrating Agent Guards to a Subjective, Fatiguing HumanEN
- 08Automatic iPhone reboot: Forensics firm allegedly stops timeEN
- 09Design Systems for AI AgentsEN
- 10Give Your AI Agent Its Own IdentityEN
- 11Launching: Nautilus - Run a coding agent from your phone, merge on your PCEN
- 12Aura - a self-hosted AI agent with temporal graph memory, written in GoEN
- 13Pi-based coding agent with hard budget capsEN
- 14Chalkboard and Avatar for AI AgentsEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.