FIELD NOTE / INSTAGRAM
The agent cannot grade its own click path.
The short film, the complete written thought, and the evidence behind it.
The Instagram edition will be linked here after its public post is verified.
The agent cannot grade its own click path.
Day 02 · 2026-09-29 · Instagram
Short video caption
Use separate actor and oracle lanes. A real success banner can belong to the wrong account; fix the target ID before execution and verify that exact record afterward. Illustrative case. #EricFieldNotes
Full written post / accessible read
A browser agent is excellent at traversing a messy form. Its trace is also biased toward what it could see. If the same session decides success from a green banner, the test may never inspect the object that was actually saved.
Imagine two customer accounts with similar names. The agent edits one and captures a success banner. The request was for the other. The screenshot is perfectly real, but the business action is wrong.
Lane one is the agent's actions and screenshots. Lane two is a fresh observer with the original request: target account ID, expected field, and final persisted state. The lanes meet only at the verdict. A mismatch stays visible.
Capture the exact target ID and expected postcondition before giving the task to the agent. Let it navigate freely, then compare an independent readback against that fixed contract. Do this because attractive UI evidence can be accurate and irrelevant.
#EricFieldNotes
Evidence and boundary
On-screen boundary: ILLUSTRATIVE UI HARNESS. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.