Rebuilding companies around self-improving AI loops

Summary

A 13-minute Y Combinator batch talk arguing that company hierarchies are obsolete once an organisation makes its domain knowledge legible to AI. Blomfield contends that AI adoption should not mean adding copilots to existing workflows; instead, redesign the company as recursive, self-improving loops that improve even while humans sleep. He walks through the five-layer loop structure (sensors, policy, tools, quality gate, learning) and gives a live example: YC built a monitoring agent that watches queries, diagnoses failures, proposes fixes to tools and context, commits code, and deploys overnight. Watch the full talk via the original video.

Key Claims

• Hierarchical Roman-legion companies designed for human information flow become redundant when you can make organisational knowledge legible to AI; loops that self-improve transcend the coordination problem.
• To enable AI loops, record everything the organisation does—office hours, Slack, decisions, telemetry—then diarize, aggregate, and synthesise it into context the models can use.
• A self-improving loop has five layers: sensors (data in), policy (what the AI can do without asking), tools (deterministic APIs), quality gate (checks and human review for high-risk actions), and learning (feedback and iteration); when minimal human intervention is required, the system compounds.
• YC's query-monitoring agent diagnoses failed queries overnight, proposes updates to tools or context, commits code, and deploys; the next morning, the same query succeeds. That is not productivity gain; it is self-improvement.
• Token budget, not headcount, becomes the constraint; YC companies are reaching 5x revenue per employee compared to 18 months ago; the next constraint will be token usage.
• Data and comprehension are permanent and valuable; treat software as ephemeral and regenerable. The real asset is domain knowledge (how the function works), not the code.
• Middle management as a coordination layer is over; every human must be an individual contributor; AI handles routing and decisioning.

Quotes

• "If it is recorded, it happened to the AI. If it did not get recorded, it did not happen."
• "Burn tokens, not headcount."
• "The valuable part is the comprehension inside people's heads—how the function works—not the software itself."
• "Human beings reach into places the models cannot go yet—novel situations, ethical considerations, and high-stakes moments."