If we treat sustainable pace and cognitive load as first-class constraints in agent-first workflows—on par with the craft bar and ambition frontier—how do our current recommendations (triad pairing, learning-PR quotas, erosion escalations, designer-owned harnesses) change in practice, and what failure modes emerge when teams optimize for judgment leverage and ambition expansion but leave dopamine-heavy agent loops and review fatigue essentially unmanaged?

dhh-agent-first-software-craft | Updated at

Answer

Treating sustainable pace and cognitive load as first-class forces small but real changes to current patterns and exposes new failure modes if ignored.

  1. How practices change in practice

  • Triad pairing (senior + junior + agent)

    • Before: maximize exposure and throughput.
    • With pace constraint:
      • Cap triad blocks (e.g., 90–120 min) and require breaks / solo consolidation time.
      • Rotate who "drives" (junior vs senior) to avoid constant high-focus for one person.
      • Limit concurrent agent tasks per triad to 1–2 to keep context stable.

  • Learning-PR quotas

    • Before: N learning PRs per week to force practice.
    • With pace constraint:
      • Track average review time per learning PR; cap total review minutes per reviewer.
      • Prefer fewer, deeper learning PRs with explicit reflection notes over many shallow ones.
      • Allow "observed" learning (shadowing, design writeups) to substitute for some PRs.

  • Erosion escalations (Arch Drift tickets)

    • Before: trigger whenever metrics cross thresholds.
    • With pace constraint:
      • Batch escalations on a cadence (e.g., weekly) instead of interrupt-driven pings.
      • Timebox response (e.g., 60–90 min per boundary per week max).
      • Maintain a small backlog and drop low-severity items instead of endless cleanups.

  • Designer-owned harnesses

    • Before: optimize for more flows and ambition.
    • With pace constraint:
      • Add WIP limits: max active experimental flows per team.
      • Require an explicit "maintenance owner" per harness surface so ops load is clear.
      • Schedule harness change windows instead of ad-hoc tweaks all week.

  • Agent loop ergonomics

    • Encourage sidecar, diff-first loops with:
      • Small diff caps per run.
      • Natural review boundaries (one ticket → one or few diffs).
      • Fast ways to pause or park in-flight agent work.
  1. New failure modes when dopamine loops and review fatigue are unmanaged

  • Hidden overwork via micro-spikes

    • Agents make it easy to chain "one more quick change"; days stretch, breaks disappear.
    • People feel constantly behind on reviews and verification.

  • Degraded judgment under review load

    • Seniors skim diffs, rubber-stamp, or default to trust in agents.
    • Real review bottleneck shifts from time to attention quality.

  • Incentive drift toward throughput

    • Metrics celebrate tickets closed and agent tasks run, not well-chosen work.
    • Juniors optimize for speed; taste and system sense stagnate.

  • Harness and prompt sprawl

    • Designers and engineers keep adding tools and flows with no pruning.
    • Cognitive load of "which tool/how to verify" climbs; new people drown.

  • Emotional dependency on agent dopamine

    • Work feels flat without constant visible progress from agents.
    • Harder tasks (spec clarification, deep debugging) are avoided.

Overall: to honor sustainable pace, teams need explicit caps (time, WIP, review load), rotation of high-focus roles, and pruning rituals for harnesses and metrics, not just better prompts.