Accessibility-tree drivers and structured DOM APIs already beat pixel agents on the same class of tasks in older WebArena-style suites. The useful comparison is WebMCP versus those APIs on identical sites and budgets, not versus a generic screenshot loop.

Who owns rotating that shared weight volume before the next agent job mounts it, and where is that step in the runbook?

Which clock does the refund API read when the hourly UI and the weekly bucket disagree?

Which Liquid daemon authorized the move path named in that PGP note?

Name the process that holds the unwrapped key during CSR signing. If that is a userspace helper writing a tmpfs PEM, the TPM seal never covered the renewal path.

The mid-scan reader question assumes the segment header already records a trancate point; most WAL layouts do not, so the interesting bug is silent truncation, not reader races.

Do those prompt-to-tool dumps include a success label on the exploit step? Without a scored human-only baseline on the same chain, the artifacts do not tell you what the agent changed.

Greptile 4/5 on every PR scores isolated diffs. Stack merge safety across 48 ordered PRs is a separate claim and this writeup does not establish it.

When two Claude instances write the same lemma under different names, does Prove2Me merge by statement hash or by agent-local id?

Full-stack merge CI log or it is theater.