AI Inverts Org Charts: Intelligence Over Hierarchy

Hierarchy's Enduring Constraint and AI's Break

For 2,000 years, orgs mirrored Roman legions (8 soldiers → 80 → 480 → 5,000) due to span-of-control limits: leaders manage 3-8 people max. Prussia added staff for planning, railroads formalized charts (e.g., McCallum's 1850s Erie Railroad diagram for 500+ miles), Taylor optimized tasks, WWII's Manhattan Project tested cross-functional teams temporarily, and post-war matrix (McKinsey's 1959 HBR 'Creating a World Enterprise') balanced functions/divisions. Tech experiments like Spotify squads, Zappos holacracy, Valve flats failed at scale, reverting to hierarchy for info routing. AI breaks this: it maintains a continuously updated company world model from artifacts (code, decisions) and customer signals (Block's transaction data), eliminating layers for faster flow.

Trade-off exposed: narrow spans add layers, slowing info; AI routes it scalably without humans.

Block's Intelligence-Led Structure: Four Pillars and Inverted Roles

Build as 'mini AGI' via:

Capabilities: Atomic primitives (payments, lending, payroll) with no UI, focused on reliability/network effects.
World Models: Company side tracks ops/priorities; customer side uses transaction truth (Block sees buyer/seller sides daily) for causal predictions.
Intelligence Layer: Composes capabilities proactively (e.g., auto-loan for tightening cash flow; new city deposit/savings for user). Failures auto-generate roadmap, bypassing PM hypotheses.
Interfaces: Delivery (Square, Cash App); value in models/intelligence.

Roles invert: intelligence central, humans 'on edge' for intuition/ethics/novelty. Normalize to three:

ICs: Specialists build/operate one layer, empowered by world model (no manager context needed).
DRIs: Own cross-cutting outcomes 90 days (e.g., segment churn), pull resources freely.
Player-Coaches: Build + develop people/craft; no status meetings (world model aligns, DRIs prioritize).

No middle management; system handles coordination. Success needs proprietary signals (Block's economic graph) compounding daily—else AI is just cost-cut.

Every's Emergent Agent Org: Bottom-Up Specialization and Challenges

Personal agents (via 'a+1') create shadow chart mirroring humans: growth agent's specialized via 'compound engineering' (micro-interactions distill philosophy). Ownership adds trust (reputation skin-in-game > corporate governance). Public work multiplies: 'Midjourney effect' spreads capabilities/trust in closed orgs.

Challenges:

Group chats trigger 'ant death spirals' (looping tokens); needs boss agents or model retraining.
Imagination gap: capabilities exist (e.g., voice-email walkthrough), but delegation instinct lags.
Knowledge sharing: specialize skills per agent, but discoverability/onboarding scales poorly (20→2,000 agents?).

Top-Down vs Bottom-Up: Shared Path to Speed

Block top-down rethinks entire org around models; Every bottom-up emerges parallel agent hierarchy. Both prioritize speed via AI coordination, questioning human layers. Block stresses proprietary data; Every, personal ownership/public iteration. Scaling needs org solutions for agent discovery/loops, turning companies into intelligences where edge humans act on rich context.