FIELD NOTE / LINKEDIN
The worker's tests are useful but interested.
The short film, the complete written thought, and the evidence behind it.
The LinkedIn edition will be linked here after its public post is verified.
The worker's tests are useful but interested.
Day 06 · 2026-10-03 · LinkedIn
Short video caption
Agent-written tests improve feedback, but can share the patch's misunderstanding. Keep consequential contracts outside the worker's writable tree; CI checks independent state and a deliberate mutant. Missing evidence is UNVERIFIED. #EricFieldNotes
Full written post / accessible read
Let coding agents write implementation tests; that feedback is valuable. But a worker that writes the patch, chooses its test cases and summarizes the verdict can produce a convincing green result from one mistaken interpretation of the requirement.
Picture a feature that renders a confirmation page correctly but commits the wrong customer tier. The worker's test asserts the page; its summary reports success. No one checked the persisted tier or the original commercial promise.
Before implementation, a reviewer defines the high-cost invariant and keeps its acceptance fixture outside the worker's writable tree. CI tests the candidate revision against that fixture and verifies persisted state. A deliberate mutant proves the fixture is sensitive to the error class.
Let the worker own its exploratory tests; let a protected runner own the consequential verdict. Block merge if the runner is skipped or its fault control survives. Do this because a test count measures activity, while an independent exam measures the promised outcome.
#EricFieldNotes
Evidence and boundary
On-screen boundary: EVALUATION ARCHITECTURE. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.