The verification tax is the real agent bill — prove once, stop green-on-green
Sep 15 BlueLupin census: code was ~8% of model output; review/verify ~40%; 15/39 Playwright runs were already-green reruns — cut avoidable proof without deleting the toll booth.
Manish Gupta’s Sep 15 BlueLupin field report measures what Anthropic’s CI strain implies on a single programme: generation got cheap; the stamp did not. On one busy session, browser acceptance alone burned ~two hours across 39 Playwright runs; 15 of those (~38% of runs, ~45% of browser wall time) re-tested scenarios that were already green because merging main voided every branch record. Sign-in ate ~42% of each run. Review and verify stages consumed ~40% of workflow model output; writing code was ~8%. Hundreds of main-thread commands were just polling a gate and re-reading a fat context.
That stack is the verification tax: CI/acceptance that re-runs untouched work, human/agent review that re-reads the same world five times, and session economics that look like “AI cost” but are waiting cost. Gupta cites Malhotra’s Anthropic 25× CI post as industry context, then ranks levers by measured saving before buying boxes. Lane A (wall time): void a browser record only when this branch changed a user-facing file (intersection of diff(testedSHA, HEAD) with the branch’s own merge-base diff — stop green-on-green); fail-fast missing acceptance before the long static gate; sign in once per persona per run; parallel reads with exclusive writes and waited= instrumentation; move saturating CI off the laptop only after the census shows load kills, not before. Lane B (tokens): one shared change-manifest brief for every review lens; drop lenses the gate already ratchets; tier models (top-tier for build/security/correctness; mid/fast for assembly and verify); per-stage effort instead of inherited maximum; one monitor, one verdict line — zero main-thread polling.
What they refuse to cut is the point: security and correctness stay on a serious model; browser acceptance stays with a register that tells the truth; a final human on the merge (the person who opened the agent session does not approve their own PR). The tax they cut is repeating proof; the tax they keep is first proof that matters.
Why it matters for code review
Continuous factories under sustained PR volume win or lose on whether verification is a product — impact-aware, instrumented, hostile to avoidable reruns — or an infinite fixed cost of every generated line. Overnight candidates only shrink morning review debt when the stamp process stops paying the toll twice for the same road.