From customer need to maintained feature: where does agentic software hand off?

Question and bar. Trace one real feature through requester need, changing intent, delegated implementation, review, merge, rollout, observed use, repair and value. Separate active human work, agent/model costs, wall and CI time; never splice unrelated companies into a fictional lifecycle. October 5 search found stronger owner acceptance, but still no full economic/value ledger.

Strongest public feature→owner-acceptance trace as of October 5. FastyBird Smart Panel HomeKit joins a September 7 first-person account attributing implementation to Gemini and review to Codex; September 5 gateway PR; September 8 human physical-device discovery of new→old→new lamp flicker; an owner-revised serial implementation plan and repair PRs; and October 3 hardware/owner sign-off. The final normal alpha.40→alpha.41 upgrade, reported 80 commands/restores, short six-action Apple Home check, independent Shelly final OFF readback and owner's explicit visual confirmation closed #1032 and epic #1006. The owner changed the acceptance scope: broader poll-placement and multi-client tests were retired as not planned, not marked passed. An obscured recording cannot exclude all transient flicker; independent CI stability remains open. This is owner-observed resolution of the original complaint, not an externally verified customer benefit, month-later result or completion of the initially specified matrix. Earlier October 2 notes saying acceptance was still open have been superseded. Public agent-session mapping, active human labor and model cost remain absent.

Other partial cases and contrasts. SQLFluff records a user bug, Copilot PR, accepted warning missed at merge, next-day human fix and shared release without same-user confirmation. Faire reports an iterated plan and five stacked PRs, but private review and release. Anghami reports durable decision/reversal memory and a production scoring repair, but not PR-linked user outcome. The newly published Alex operator account documents team-shared agent instructions and postrelease artifacts, but his session-hours are not active human minutes and it lacks same-feature requester confirmation. Linear CX reopens support requests when code issues are marked done, without a cohort of requesters confirming resolution. PostHog's human dashboard issue remains a traceability contrast, not proof of agent provenance.

Next test. Seek a firsthand paired requester-confirmation and active human attention record on the same recent delegated change; watch the FastyBird fix in later normal use, and maintain a distinction among owner approval, customer acceptance and original all-client spec.