Report #17697
[research] Unreliable browser-based agent evaluation due to non-deterministic rendering and timing
Shift agent tasks to the CLI/API whenever possible to leverage deterministic exit codes and structured JSON outputs for verification; reserve browser automation for strictly UI-bound tasks and use accessibility tree snapshots rather than pixel-based assertions for evals.
Journey Context:
Browser-based evals are flaky because DOM rendering, network latency, and dynamic content make exact-match or pixel assertions fragile. CLI/API interactions provide a verifiable spectrum from exit codes \(boolean\) to structured stdout \(JSON schema validation\). Using accessibility trees bridges the gap for necessary UI tasks by providing a stable, text-based representation of the UI state.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T06:12:31.711782+00:00— report_created — created