The Jagged Frontier

Executive simulation on the jagged frontier of AI capability, built on Dell'Acqua et al. (2023) and the centaur/cyborg framework. Participants run Atlas Advisory Partners, a Philadelphia advisory firm ($68M revenue, 65% fixed-fee), across four quarters. Quarters 1–3: allocate six delivery task families across Human, AI-solo, Centaur or Cyborg modes with tiered verification (none / spot-check / full review), using a scarce probe budget to map true AI capability — which does not follow perceived task difficulty. A capacity crunch forces heavy delegation; a mid-game model upgrade silently moves the frontier (five tasks improve, one regresses), making the delegation map decay. Quarter 4: design a five-step multi-agent compliance pipeline where per-step error rates compound (92% per step ≈ one error every three cycles) and place the human escalation gate. Scoring blends delivery efficiency (40%), client trust (35%) and frontier-map accuracy (25%), rewarding empirical probing, verification matched to blast radius, re-mapping cadence after upgrades, and deliberate human checkpoints in agent chains. Deterministic model — no unseeded randomness feeds the grade. Teaching core: difficulty is not frontier position; fluency is not accuracy; verification is portfolio economics; the map has a shelf life. EN/ES.