Webcmd Saves Website Routes for Repeat AI Agent Tasks

A browser agent can finish a task on a website, then return and spend time rediscovering the same menus, routes, and limits. Webcmd addresses that repeat-visit problem by saving what an agent learns about a site for later use. Its value lies in reducing exploration, not in taking over the agent’s decisions or guaranteeing that a saved route will always work.

A site map built during the first task

Webcmd works as a skill for agent tools including OpenAI Codex, Claude Code, Hermes, and OpenCode. It can work with existing desktop browsers such as Google Chrome and Brave, rather than requiring users to move to a separate browser. As an agent navigates, Webcmd builds persistent memory for the website’s domain: useful paths, page structures, shortcuts, request constraints, rate limits, and warnings.

That first visit still requires exploration. In a Reddit research task, an agent was asked to find five recent posts about AI browser automation and return each post’s title, subreddit, short summary, and URL. The instruction specified that the task was read-only. The agent had to find workable ways through the site before it could gather the posts.

▲ Domain routes and constraints

The resulting memory included a useful distinction between routes: Reddit’s older interface or an RSS feed could provide a cleaner path for this research than its newer desktop pages. Webcmd saved that kind of site-specific guidance so a later task would not have to discover it again. The stored knowledge concerns how to navigate and retrieve material; the agent must still find content relevant to the current request.

Reuse without blindly trusting old paths

On the second Reddit run, the agent used its saved navigation profile instead of exploring the page structure from scratch. Webcmd supplied routes and constraints, while the underlying model retained responsibility for choosing actions and judging the results. That distinction matters: memory can narrow the search for a workable path, but it does not replace reasoning about the task.

Saved routes can become stale when a website changes. A selector—an identifier an agent uses to locate a page element—may no longer point to the right place after a layout update. Webcmd checks paths against the current document object model, or DOM, the structure of the page. When it finds a mismatch, it can revise its stored knowledge and remove an obsolete route rather than continuing to rely on it.

▲ Revalidated browser path

The repeated task illustrates the intended workflow, but it does not supply a measured time or token saving for that Reddit example. The practical benefit depends on whether the agent can reuse the recorded guidance for the site and task at hand, and whether changed pages can be checked and updated successfully.

What the benchmark measures

Webcmd was also evaluated on BU-Bench V1, a 100-task browser automation benchmark. The evaluation used Codex GPT-5.4 as the judge and CloakBrowser as the underlying browser engine. In that comparison, Webcmd had the highest task accuracy, the lowest estimated controller API cost per completed task, and the fewest average agent turns per task among the listed tools.

Those rankings describe BU-Bench V1, not every website or agent setup. The project’s repository separately advertises up to 10 times lower token spending on repeat executions; the benchmark table reports controller API costs, not a token-saving measurement for the Reddit task. Keeping those two kinds of evidence separate makes the benefit easier to assess.

How to use the approach

Webcmd is open source under the Apache 2.0 license. For developers considering it for repeated browser work, the workflow is straightforward:

1. Install the package with npm install -g @agentrhq/webcmd, then add its skill with webcmd skills add in a supported agent setup.
2. Let an initial run explore a site so Webcmd can record useful routes and constraints. For research tasks, explicitly instruct the agent to remain read-only so posting is outside the requested work.
3. Reuse the saved domain memory on later tasks. For concurrent agents, use Profiles and Spaces to separate browser sessions and execution contexts, including cases involving different account logins.

Profiles and Spaces address a different problem from route memory. They keep simultaneous agents from competing for window focus or interfering with one another’s sessions; they do not determine whether an individual agent has made the right decision.

The takeaway

Webcmd’s main idea is to preserve a website’s working routes while checking that they remain valid. The Reddit task shows how that can spare a returning agent from some exploration, and BU-Bench V1 provides a defined comparison of accuracy, cost, and turns. If repeat visits are a substantial part of a browser-agent workflow, start with a read-only task, let the first run build a map, and check whether later runs actually use and update it.