AI Agents Flatten Hierarchies with World Models
AI replaces human info-routing in org charts via company/customer world models and intelligence layers, enabling edge-focused roles like ICs, DRIs, and player-coaches for faster coordination.
Hierarchy's Info-Routing Limit Exposed by AI
Traditional org charts, from Roman legions (8 soldiers per decanus, scaling to 80/480/5,000) to railroads and Taylorist pyramids, solve coordination via nested spans of control (3-8 people per leader). Prussia added staff for planning, military brought line/staff distinctions to business, and post-WWII matrix models (e.g., McKinsey 7S) balanced functions with agility. Yet experiments like Spotify squads, Zappos holacracy, and Valve flats fail at scale due to slow info flow—more layers delay decisions. AI breaks this: agents maintain a continuously updated company world model (from remote artifacts like code/decisions) and customer model (from transaction data), eliminating human relays. Block leverages millions of daily Cash App/Square transactions for causal predictions, turning honest signals (spend/save/send) into proactive intel.
Block's Intelligence Layer Inverts the Org
Build four layers over models: (1) Capabilities (payments/lending primitives, no UI, network-protected); (2) Dual world models (internal ops + per-customer reality); (3) Intelligence composes solutions proactively—e.g., auto-loan for tightening restaurant cash flow or city-move savings for users—without PM roadmaps; (4) Interfaces (Square/Cash App) deliver. Failures auto-generate backlogs from customer reality. Humans shift to edge roles: ICs build/operate layers with model context (no manager needed); DRIs own 90-day cross-problems with resource pull authority; player-coaches code + develop people, skipping status meetings. Result: no middle management, system handles alignment/priorities. Deep proprietary understanding (Block's economic graph) compounds advantage; without it, AI is mere cost-cut.
Emergent Agent Orgs Amplify from Bottom-Up
At Every, personal agents (e.g., Montaigne for growth, R2C2 for product) form shadow org charts mirroring human specializations via compounded interactions—'compound engineering' distills philosophy without manual docs. Ownership adds trust: your agent stakes your rep, unlike shared Claude. Public work creates 'Midjourney effect'—watching agents handle MRR or bugs teaches org capabilities/trust in closed teams. Challenges: agents flop in group chats (ant death spirals burn tokens; needs boss agent or model retraining); imagination lags tech (e.g., voice-email walkthrough sat unused until need struck). Parallel agent layer speeds ops, raising how-work-gets-done questions.