A saved navigation decision, retrieved by a fresh agent

A fresh agent, given only the public agent guide and the name of this site’s Module, retrieved our saved navigation decision from Cortex and checked the header of eduarda.ai against it on desktop and phone. Nobody restated the rule. The check and its written receipt took under seven minutes.

Checked October 17, 2026. This is a public excerpt of the recorded navigation check, not the private knowledge graph or a controlled performance study. The story is Meet AI eduarda.

Task and source

The fresh agent was asked to use the public agent guide, retrieve the website Module and check homepage navigation at desktop and phone widths. It did not receive the parent conversation or the saved rule in its prompt. It used an existing connected Codex session and retrieved eduarda-web version 61 through Cortex Flash, including its dependencies.

Saved decision excerpt

Hide the complete header container at the shared desktop-sidebar breakpoint, including reserved height, glass and backdrop effects; hiding only children left the empty strip human Eduarda rejected.

The same decision retained the phone menu, home portrait and theme control. Desktop already has navigation in its persistent sidebar; the phone needs its header controls.

Fresh agent's recorded observations

Desktop, 1440 × 900:

Rendered DOM measurement: the header had display:none and a 0 x 0 rectangle. The main region began at x=256, y=0.

Phone, 390 × 844:

Opening the menu exposed the same complete grouped destination inventory as desktop, with Start here selected and Set up Cortex last.

Opening moved focus to the product switcher inside the sidebar. Escape closed the drawer and restored focus to Open menu.

The agent also checked a direct pointer opening/dismissal, destination selection, and the mobile header hiding on downward reading and returning on upward scrolling. No human restated the rule or intervened during the task. Verification and cleanup took 5 minutes 42 seconds; the checked written receipt was complete after 6 minutes 42 seconds. This is elapsed time for one task, without a comparison baseline.

Limits

This was a light-theme homepage navigation check in one browser, using an already-connected agent. It did not test a new installation, provider switching, implementation by a fresh agent, every page, or all devices. One locator click moved the reading position; a direct pointer check preserved it. The recorded result distinguishes that automation discrepancy from site behavior.

The agent did not open the prior conversation or local audit files. Broad Cortex discovery returned incidental unrelated Module results; they were not used as navigation evidence. The trial therefore establishes retrieval and application in this bounded task, not absolute absence of incidental context exposure.

Saving knowledge and observing a later correct result do not by themselves prove causation, a speed improvement, or exponential self-improvement.