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

Question and bar (September 28, 2026). Follow one real feature through customer need, human-owned intent, plan and revision, several delegated implementations, operator orientation, review decisions, release and follow-up. A strong answer records who intervened when, active human and agent costs, product outcome and repair. Do not stitch distinct vendor reports into a fictional pipeline.

Four subquestions. (1) How are criteria, dependencies and plan revisions visible during multi-PR work? (2) Who owns product decisions, risky code, rollout and CX follow-up when non-engineers start the work? (3) What lets a human reviewer judge fit rather than merely parse code—and when may human diff review be waived? (4) Which measurements join agent spend, active human supervision, accepted change, customer use and 30-day repair?

First-person and customer-operating sources read. Linear’s composite customer email → CX scoping → PM synthesis → human-assigned engineer → agent/human review → merge/Done → CX notification shows named owners, but not a single multi-PR feature history; a PM-initiated filter fix is one small worked example. Linear reviewer Maciek Pekala first retrieves the originating problem, tries a preview and checks system/UX fit; claims quality confidence, not verified comparative defects. Linear’s auto-fix engineer narrowed an unsuccessful all-bugs approach to task-shaped gateable changes; human keeps harder ambiguity. OpenAI Symphony makes issue status, human-approved plans, dependency trees and video review packets visible, but loses interactive mid-flight steering. Cursor Projects instead keeps a project coordinator and shared context through parallel work and after shipping. Amplitude’s Cursor case links Slack customer requests, Linear ticket, agent PR and risk-based reviewer routing, but has no recorded customer case or full release metrics. These are distinct systems and accounts.

Quality and role checks. Grab designers’ production UI fixes show non-engineer implementation with design/engineering checks, not general role interchangeability. Linear’s human-written human-reviewed permission incident shows why high-blast-radius paths need explicit test/rollout controls regardless of authorship; do not label it agent-induced. Review-autonomy map, economics map and follow-up-fixes paper supply evidence lenses, not missing field measurements.

State of evidence / next query. No source read yet establishes an individual multi-PR feature with evolving spec, exact human-time ledger, release controls, customer usage and future fixes. Next seek original issue/PR/incident links or permission to examine a real internal trace, then look for cross-functional marketing, sales or support-owned changes with explicit risk gates and product consequences.