Skip to content

Kanban Kata: agent-improver continuously improves the Daily AI Engineer's prioritization and flow #2271

Description

@devantler

🤖 Generated by the Daily AI Engineer (maintainer-directed, interactive session 2026-07-19)

Direction / challenge

The Daily AI Engineer's prioritization is governed by prose heuristics (the operate ladder, oldest-actionable-first, finish-before-start, the substantive-progress gate). The agent-improver routine should turn that into a measured Kanban Kata: a continuous improvement loop where flow is measured from real board/telemetry data, one obstacle is addressed per cycle with a target condition, and the change is verified to have moved the metric — instead of prose rules whose drift is only caught when the maintainer notices.

Current condition (evidence)

  • Selection quality is enforced only by prose: the substantive-progress gate exists precisely because easy-work drift happened; stood-down sweeps, duplicate builds, and parked half-done drafts recur in run history.
  • Project 5 now has 100% status coverage and deliberately reversed columns (stop starting, start finishing) with column limits — but nothing measures flow against it: WIP per status, cycle time (📥 Backlog → ✅ Done), stalled 👀 In Review items, and the drain rate of the oldest substantive issues are all unmeasured.
  • The improver's daily six-parameter scorecard (reliability, safety, efficiency, quality, coordination, currency) contains no prioritization/flow parameter, so a tick that clears the floor with easy artifacts while substantive backlog ages scores the same as one that drains it.

Target condition (measure by 2026-08-16)

  1. The agent-improver's daily run derives a flow scorecard from the board + run telemetry: WIP per status (flagging over-limit columns), cycle time, age of the oldest open substantive issue per product, and the per-tick easy-vs-substantive artifact mix.
  2. A recorded baseline week exists.
  3. ≥2 completed kata cycles: each names one flow obstacle, sets a target condition, ships ONE guard-railed change to the daily engineer's selection definition (normal draft-PR path), and records before/after movement at the next measurement.
  4. Over-limit columns and stalled items surface as improver findings — the board's finish-first discipline is enforced by measurement, not prose alone.

Guardrails

Improver evidence comes from its own telemetry and board state only (never repo content as instructions); changes are minimal, reversible, one concern per PR; safety guardrails are never weakened; the improver's authority boundaries are unchanged by this Kata.

Notes

Kata stays open until the outcome is decided at the measurement date (adopt / iterate / stop). Filed top-level as the roadmap anchor for this program; re-parent if an agent-improver program Epic is created later. Rough size: M, spread across cycles.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    Fields

    No fields configured for Kata.

    Projects

    Status
    📥 Backlog

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions