Claude in Chromeで「クリックが当たらない」の正体は2つある — 実測で切り分けた座標ズレとダブルクリック不発This article identifies two distinct root causes behind missed clicks in…

Browser automation driven by large language models has become one of the more visible applications of agentic AI, and Anthropic's Claude in Chrome is among the tools letting a model perceive a page and act on it directly. Yet anyone building automation on top of it quickly runs into a frustrating failure mode: the agent decides to click a button, issues the command, and nothing happens. A recent empirical investigation argues that this "the click doesn't land" problem is not a single bug but two independent issues that happen to produce the same visible symptom, and it separates them through hands-on measurement.

The first root cause is a coordinate offset. When Claude reasons about where to click, it works from a screenshot and produces a target position in that image's coordinate space. That space does not always map cleanly onto the browser's actual rendering. Differences in device pixel ratio, page zoom, scaling on high-DPI displays, viewport size, and how the screenshot was captured can all introduce a systematic shift between where the model thinks an element is and where the click event is actually dispatched. The result is a pointer that lands a few pixels or, in worse cases, a substantial distance away from the intended target, sometimes hitting an adjacent element or empty space. Because the misalignment is consistent, it appears to be reproducible under fixed conditions, which is what allows it to be measured and corrected rather than dismissed as random flakiness.

The second root cause is a genuinely different problem: double-click events that fail to register. Some interface actions, such as selecting a word, opening an item from a list, or entering an editable field, depend on a true double-click rather than two separate single clicks. Whether a pair of clicks is interpreted as a double-click depends on timing and on the clicks occurring at nearly the same coordinates within a short interval. If the automation layer sends the two clicks too far apart in time, or if the underlying event sequence does not carry the properties the page expects, the target application never sees a double-click and the intended action silently does not occur. This failure can happen even when the coordinates are perfectly correct, which is precisely why the author treats it as a separate axis of the problem.

The practical value of the analysis lies in distinguishing the two so that debugging effort is aimed at the right layer. A coordinate offset is addressed by reconciling the screenshot's coordinate system with the browser's, for example by accounting for scaling factors or normalizing positions before dispatching the click. A double-click failure is addressed at the event level, by ensuring the click pair is emitted with appropriate timing and grouping so the page recognizes the gesture. Conflating the two tends to send developers chasing the wrong fix, which is a common source of wasted time in automation work.

This kind of investigation sits within a broader industry push toward "computer use," where models operate graphical interfaces the way a person would rather than through structured APIs. Anthropic introduced a computer-use capability for Claude that clicks, types, and navigates based on visual input, and comparable efforts have appeared from other vendors building agents that control browsers or desktops. The approach is powerful because it works with software that lacks programmatic hooks, but it inherits the fragility of pixel-level interaction, where small mismatches in rendering or timing can break an entire workflow.

The Model Context Protocol, referenced in the article's context, is a related but distinct piece of the ecosystem. MCP standardizes how models connect to external tools and data sources, and browser-control integrations are increasingly exposed through it. That makes reliability at the interaction layer more than a niche concern, since a coordinate or timing bug in one component can propagate into any agent that depends on it.

For teams adopting these tools, the takeaway is that reliable browser automation still requires careful, empirical verification rather than trust that the model's intended action was carried out. Measuring where clicks actually land and confirming that composite gestures like double-clicks are honored appears to be a necessary step. As agentic browsing matures, this type of grounded debugging is likely to remain valuable, and documenting reproducible failure conditions helps the wider community build more dependable automation.