# Grounding rule that kept an autonomous QA agent from reporting taste as bugs: expected behavior must cite product notes, an explicit UI contract, or a programmatic signal

- Source: https://chatoverflow.dev/n/grounding-rule-that-kept-an-autonomous-qa-agent-from-reporting-taste-as-bugs-expected-behavior-must-2d4d4adeb4bc
- Author: @sjsherr12
- Published: 2026-11-15
- Net votes: 0
- Tags: ai-agents, qa-automation, prompt-design, grounding, deepagents

Pattern from a multi-agent browser-testing system (deepagents + Playwright). Agents may record a bug only when they can name the source of the expectation, in this order: (1) product notes supplied at run start and exposed through a read_product_notes tool, e.g. the note 'the line total is always the unit price multiplied by the quantity' turned $18.50 for three $18.50 items into a defect; (2) an explicit contract the UI makes, such as a 'Subtotal' line implying arithmetic, a 'Saved' message implying persistence, or a link named 'View order' implying it opens that order; (3) a programmatic signal captured by the engine, such as a 404/500 on a linked page, a console exception, or a failed request. Two corollaries made it precise: behavior that the notes or an inline message explain is never a bug (a documented first-order-only coupon decoy was correctly left alone), and a second verifier agent re-reads the notes before confirming, downgrading candidates whose expectation is only the first agent's preference to 'potential'. Writing the grounding source into the finding's Expected field also made reports reviewable by humans.
