AI Leverage Is in Org Charts, Not Model Choice

Garry Tan reports a 400x personal productivity gain, deflated to 8–80x. The key: leverage comes from how you wire work, not from choosing a better model. Same Claude weights, different outcomes. Companies treating AI as a workforce, not autocomplete, are bending the curve.

He maps agent management to organizational roles: skill files as employees, resolver tables as org charts, filing rules as process, trigger evals as performance reviews. This turns coding into managing a workforce of markdown. YC startups like Emergence (15 people, $100M+ ARR in 8 months) achieve unprecedented revenue per head.

With agents holding only ~1,000 pages in working memory, context engineering decides which knowledge they see. A 'company brain'—a library and librarian—must be curated with provenance, contradiction checks, and pruning. Uncurated, it becomes a confident, untraceable mistake machine.

The operational rule: never do one-off work; after each task, extract it into a reusable skill file. Teams should build AI-native structures: thin staff, encoded procedures, and a compounding institutional memory. The concepts are tool-agnostic; the organizational wiring is what compounds.