Can a delegated software change pair human attention with requester-verified closure?

Research brief and result, October 6, 2026. Question: can a recent primary account connect a real request, a coding-agent's revisions, the human's active planning/review/test time, and the same requester's later observed outcome? Good answer identifies the request cohort and dated release, counts human minutes separately from elapsed agent sessions, records reopenings or same-issue return, and retains failed attempts. Earlier October search found FastyBird owner acceptance without an independent requester or active labor minutes, and Alex session-hours without active human time.

Four subquestions searched: (1) recent original engineer issue→PR→customer confirmation plus supervision minutes, (2) intervention study instrumenting active review time and postrelease outcome, (3) support/internal-request systems connecting original requester to handler minutes, (4) new FastyBird repair after October 3 acceptance. Targeted follow-up queries looked for quoted 'customer confirmed' PRs and minutes in Linear and FastyBird. No publicly verifiable same-feature pairing turned up. FastyBird #1032 still records October 3 scoped owner acceptance and distinct open CI stability, not an external requester; a search for October 5 HomeKit regressions yielded no original counterexample, which does not prove none exist.

Short source list and what each tests. Qi et al.'s September 29 paper reviews 2025 agent performance PRs and re-executes selected claimed gains, showing that a merge and even a changed test can disagree with an independent measured workload. It offers an actual Devin issue/PR and later human fix, but no reviewer active minutes or original requester's follow-up. Span's first-party May–July telemetry joins prompts, human turns and reviews for 103 teams, finds observational associations with accepted code/review rework, and explicitly disclaims causality; human turns are not human minutes, accepted lines are not confirmed need. Linear's Tim Qi telemetry gives aggregate app minutes by function and more PRs opened in connected teams, but no issue-specific supervision time or customer confirmation. Klaviyo/Linear customer case describes a customer report through Zendesk→PM→Cursor→engineer→release within three business days, without published original ticket/PR, minutes, or customer return.

Synthesis and disagreement. FastyBird is better than PR-count evidence on owner-observed behavior but not independent outcome or total cost. Span offers trace-level human turns but not elapsed effort and risks conflating more accepted code with value. Linear’s aggregate in-app time rises for coordination while formal planning views barely change; it warns that PR growth does not establish business results. Qi et al. detect an independent test-oracle gap and an approval-status ambiguity but sample old open-source performance work. These are complementary missing columns, not a joined ledger or contrary causal results. Next useful work is a public owner/maintainer who times hands-on sessions and exposes post-release requester confirmation on the same ticket; absent that, do not repeat generic search. A prospective instrumentation protocol should log before/after dates, reviewer touch time, model/runtime spend, discarded attempts, customer check and return window by one request ID.