Exactly what left your machine

• Workflow density

  Every multi-agent workflow you invoked — the curve saturates around 10.

  W = 163 11 × Hill(W, 10, 1.2)
  +10.6 / 11
• The fleet (delegation)

  The total of agents you dispatched — the real fleet, and the dominant term of this pillar.

  G = 1684 12 × Hill(G, 120, 1)
  +11.2 / 12
• Concurrent peak (bonus)

  Your peak of agents running at the same instant — a hardware-capped bonus, no longer the main axis.

  Fmax is capped by your CPU at min(16, cores − 2) — it measures the concurrent peak, not your fleet. The fleet is G = 1684.

  Fmax = 13 4 × Hill(Fmax, 3, 2)
  +3.8 / 4
• Breadth of fronts

  Distinct projects you orchestrated across — conducting several fronts is part of the craft.

  proj = 2 3 × Hill(projects, 4, 1)
  +1.0 / 3

• Verify after edit

  The share of editing sessions where you ran tests, builds or re-reads afterwards.

  r = 76% ne = 37 8 × r^0.7
  +6.6 / 8
• Verification breadth

  How many distinct ways you verify — tests, builds, lint, browser — out of 4.

  mod = 3 6 × (mod ÷ 4)
  +4.5 / 6
• Human sign-off

  AskUser moments plus explicit plans — checkpoints where a person approves before the fleet moves.

  S = 52 4 × Hill(S, 8, 1)
  +3.5 / 4

• Tool diversity

  Distinct tool names used at least twice in the window.

  T = 61 6 × Hill(T, 12, 1.5)
  +5.5 / 6
• Advanced tooling

  ToolSearch and Skill invocations — reaching beyond the defaults.

  A = 202 6 × Hill(A, 10, 1)
  +5.7 / 6
• Ecosystem (MCP + plugins)

  Distinct MCP servers reached, derived server-side from the mcp__ tool names.

  E = 61 4 × Hill(E, 4, 1)
  +3.8 / 4

• Pull requests

  PRs linked in the window — shipped work, not just motion.

  P = 34 8 × Hill(P, 6, 1)
  +6.8 / 8
• Magnitude

  Distinct files touched across the window.

  M = 242 4 × Hill(M, 40, 1)
  +3.4 / 4
• Efficiency

  PRs per unit of orchestration — η = P ÷ (1 + W + G/5).

  η = 0.07 2 × Hill(η, 0.5, 1)
  +0.2 / 2

Craft starts at the full 12 pts and can only lose points — four capped penalties for messy runs.

• Compaction penalty

  Average context compactions per session — long, bloated contexts cost points.

  comp = 0.00 − min(comp × 2.5, 2.5)
  +0.0 cap 2.5
• API error penalty

  Average API errors per session.

  err = 0.43 − min(err, 2.0)
  −0.4 cap 2
• Denial penalty

  Average tool denials per session — permissions friction.

  den = 0.16 − min(den × 1.5, 1.5)
  −0.2 cap 1.5
• Interruption penalty

  Average interruptions per session — how often a run had to be stopped.

  int = 0.47 − min(int, 2.0)
  −0.5 cap 2

• Loops

  ScheduleWakeup and Monitor events — work that keeps running without you.

  L = 72 5 × Hill(L, 6, 1)
  +4.6 / 5
• AFK leverage

  Your longest run with you away — it only counts with real verification behind it.

  Gate: only scores when Verification is above 4 pts. AFK without verification is not autonomy.

  Rafk = 497 3 × Hill(Rafk, 40, 1) × gate
  +2.8 / 3
• Consistency

  Active days across the month, discounted by how concentrated the activity is (Gini).

  days = 19 Gini = 0.34 2 × (days ÷ 30) × (1 − 0.5 × Gini)
  +1.1 / 2