As designer-implementers take on more ownership of agent harness flows and reversible hunch probes, what concrete apprenticeship patterns (shadowing, paired harness design, staged autonomy on change-types) best preserve deep engineering taste and reduce apprenticeship decay for juniors who mostly interact with agents instead of hand-writing large features, and how could you tell if the seniority gap is widening or shrinking under those patterns?

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

Answer

Bullet view:

  1. Concrete apprenticeship patterns

  • Guided shadowing on flows

    • Juniors sit with a senior designer-implementer for full agent-first loops (ticket → brief → harness flow → diff → review).
    • After each run, junior writes a short “why this diff is good/bad” note.
    • Senior comments directly on that note, not just the code.

  • Paired harness design sessions

    • Weekly 60–90 min where junior and senior co-design or refactor 1–2 harness flows (CLI/MCP commands, prompts, checks).
    • Senior narrates taste: boundaries, naming, risk lanes, what not to expose to agents.
    • Junior operates the tools; senior only vetoes and explains.

  • Staged autonomy by change-type

    • Tier 0: safe UX-only and cosmetic changes. Junior owns harness flows + merges with light review.
    • Tier 1: medium-risk app behavior. Junior proposes harness usage and verification; senior must OK plan and final diff.
    • Tier 2: high-risk classes (money, auth, migrations, cross-system per 2aadec1c-*). Juniors only co-author; seniors decide.
    • Promotion between tiers gated on checklists: number of clean PRs, quality of briefs, correctness of self-assessments.

  • Diff-reading reps without agents

    • A fixed quota of “learning PRs” per week where the junior only reads and critiques diffs (their own or others’), including agent-generated ones.
    • Focus prompts: boundaries, invariants, test sufficiency, naming.

  • Intent-to-diff reviews

    • Before any non-trivial change, junior writes 5–10 line intent + harness plan (which tools, what checks, what they’re worried about).
    • Senior reviews the plan quickly; junior then runs agents and implements.
    • Post-merge, junior compares realized diff vs plan and notes gaps.

  • Rotation through verification ownership

    • Juniors own small verification scripts or scenario tests for one or two flows.
    • They must keep these green across agent changes, learning where agents break invariants.
  1. Signals the patterns are working (seniority gap shrinking)

  • Review and defect metrics

    • For tier-0/1 work, junior PRs approach senior PRs on:
      • Review cycles per PR.
      • Share of comments about fundamentals (arch/safety) vs nits.
      • Post-merge bug rate on junior-owned areas.

  • Harness and brief quality

    • Juniors’ briefs and harness changes increasingly need no rewrite of:
      • Boundaries (no cross-service leaks without reason).
      • Risk tagging and lane choice.
      • Verification steps (fewer “add tests” comments).

  • Taste articulation

    • In 3–6 month intervals, ask juniors to walk through:
      • One agent-generated diff they approved and why it fits.
      • One they rejected or heavily edited and why.
    • Compare depth and consistency of reasoning over time.

  • Autonomy mix by change-type

    • Rising share of medium-risk work that juniors can carry through with only one senior touch.
    • Stable or falling incident rate in those areas.
  1. Signals the seniority gap is widening

  • Review pattern red flags

    • Seniors rewriting large portions of junior-initiated harness flows or diffs.
    • Many “looks fine, I guess” approvals on sizeable agent diffs, followed by cleanup PRs.

  • Agent over-reliance

    • Juniors struggle to explain why a diff is safe or tasteful beyond “tests pass” and “agent suggested it.”
    • Learning work gravitating to prompt tweaks instead of understanding core code paths.

  • Drift in high-risk areas

    • Repeated boundary leakage or test gaps around auth, money, migrations, or cross-system flows.
    • Seniors quietly taking those domains back, reducing junior exposure.

  • Skill-formation stall

    • After 6–12 months, juniors still uncomfortable starting work from a textual intent and rough plan; they only feel safe when starting from existing harness templates.
  1. How to instrument this lightly

  • Tag PRs by:

    • Author seniority.
    • Change-type tier (0/1/2).
    • Agent involvement level (none / assist / heavy).

  • Track over time:

    • Review rounds and time-to-merge by tier and seniority.
    • Bug/rollback rate by PR tag.
    • Fraction of junior PRs where seniors mainly edit:
      • Plan/brief.
      • Harness flows.
      • Core code.

  • Add quarterly qualitative checks:

    • Short, structured interviews or written reviews where seniors rate each junior on:
      • Flow taste (does their harness use match team norms?).
      • Risk sense (do they lane changes correctly?).
      • Verification habit (do they propose reasonable checks unprompted?).
    • Compare cohorts over time to see if new agent-era juniors converge toward pre-agent baselines faster or slower.