Greg Isenberg's 48 Hours in SF: Data Brokers Making $2M/Week, Agent Harnesses as the New GPT Wrapper

FinanceYF5 · x · 2026-10-16

After 48 hours in San Francisco talking to billionaires, data companies, and AI founders running $1B–$100B businesses, Greg Isenberg shared 21 observations. Key points:

Data business is booming: data brokers reportedly earn $2M per week selling corporate data to AI labs; unclear if this is the start or a bubble—the interesting question is which data niches remain unclaimed.
The best question for any AI startup: "If a new model ships next month, does your company get better or worse?"—half the table went silent.
The new unicorn bar is $1B revenue in 12 months, yet few discuss retention; AI revenue arrives fast and can vanish fast.
Computers are being reinvented: model = CPU, agent harness = OS, skills/plugins = apps—useful for spotting market gaps.
Agents that learn your job by watching a week of screen recordings may be only 12 months away.
Agent harnesses are the new GPT wrapper: top teams treat models as swappable components, prefer portable prompts over fine-tuning, and evaluate on real user behavior (edit/export/share). A vertical harness = industry data hookup + domain-expert evals + permission rules + human review UI; the model may be only 20% of the product.
Software reverse-engineering/cloning is surging: someone demoed 10 open-source clones of apps each worth $5B+.

Original post →

Greg Isenberg's 48 Hours in SF: Data Brokers Making $2M/Week, Agent Harnesses as the New GPT Wrapper

More from coding & agent