Can an agent keep a backlog honest?

Hypothesis

Backlogs are where ideas go to feel remembered. A weekly agent pass that classifies every open item, pushes each through an explicit stage-gate (kill / park / promote), and verifies that "done" things actually happened should keep the backlog a decision tool instead of a graveyard.

What we built

A weekly triage routine inside the AI OS: a classifier sweeps the backlog and labels each item, a stage-gate forces an explicit state instead of an implied "someday", and a verification step samples completed items to confirm reality matches the record. Results arrive as a compact weekly summary with the contentious calls surfaced for a human decision.

Learnings

• The kill decision is the valuable one. An agent willing to recommend deletion — with a reason — does the thing humans avoid.
• Verification-of-done catches quiet drift: items marked complete whose artifact never shipped are more common than anyone admits.
• Weekly cadence is right: daily triage nags, monthly triage rots.

Log

• 2026-07-05 — Verify step live; weekly summary format settled.
• 2026-06-25 — Classifier + stage-gate armed on the OS backlog.