
Company building as agent loops

Why companies are becoming a series of loops | Anish Acharya (a16z)
Topic 3 of 19
Published Sep 6, 2026 · Clip 11:28–20:22 (8:54)
Anish traces the progression from prompts to agents (models in a loop with tools and memory) to loops (sets of agents doing tasks). Coding is the ideal loop domain because it has the best models, the most technical customers, and established processes like the five-minute bug-fix loop; his key question is what the equivalent 'business loops' are for a GM across marketing, sales, support, and legal, cascading from a loop per person up to loops running large parts of the company. He illustrates with a growth-team loop where every variant is generated, measured, and merged at statistical significance with a long-term holdout, but stresses loops plateau at local maxima and need human intuition to reach the next hill. Examples include OpenAI's go-to-market team using Codex more than engineers, a PM building everything and simulating users, and Kavak's agent-per-customer that calls a human for coaching when stuck.