How to Build a Self-Improving Company with AI
Y Combinator
YC General Partner Tom Blomfield argues that companies should treat AI as part of their operating structure, not as a copilot added to an existing hierarchy. His proposed unit is a recursive loop that observes real-world signals, applies policies, calls deterministic tools, passes quality gates, and learns from outcomes.
This is a conceptual batch talk backed by first-party YC examples, not an evaluation of self-improving companies. The recording does not provide methods for its revenue-per-employee comparison or validate the generated advice or deployed code. It does not resolve the consent, privacy, security, retention, and governance risks of recording organizational communication and automating consequential actions.
Key Points Covered
- Replace the copilot metaphor with a systems-design question: Blomfield argues that adding a more capable tool to an old hierarchy leaves the company’s structure unchanged. He instead asks founders to make domain knowledge legible enough for software and agents to act on it [00:00:00]-[00:02:24].
- Build each loop from five explicit layers: Signals such as support tickets and product telemetry feed a policy layer. Deterministic tools perform actions. Evals, safety filters, or human review form a quality gate. Observed results become the next learning input [00:02:24]-[00:04:12].
- YC’s query agent is the talk’s concrete self-improvement example: Blomfield says a monitoring agent inspects failed employee queries and identifies missing tools or data access. He says it writes a change and sends it through agent review, merge, and deployment so a later query can succeed. The talk does not show the system, its success rate, or which changes require human approval [00:04:12]-[00:05:50].
- Product and support loops remain proposals: He extends the pattern to sales-funnel experiments and customer requests that could trigger product changes without human intervention. These examples illustrate the intended architecture rather than reported production outcomes [00:05:50]-[00:06:29].
- “Burn tokens, not headcount” is a forecast, not a measured rule: Blomfield reports that YC companies reach Demo Day with roughly five times more revenue per employee than 18 months earlier. He predicts token supply will constrain growth before hiring does. No cohort definition, distribution, comparison method, or causal evidence accompanies the figure. He also acknowledges that token-usage leaderboards are gameable [00:06:29]-[00:07:23].
- The flatter-company claim is deliberately strong: He predicts that AI can replace much of middle management’s coordination role, leaving individual contributors and one named directly responsible person for each outcome. The talk does not examine larger-company evidence, management work beyond information routing, or failure cases [00:07:23]-[00:08:05].
- Legibility requires capture and compression: Blomfield recommends recording emails, Slack messages, direct messages, conversations, and office hours. He recommends reducing the material into useful context and breadcrumbs. The talk treats comprehensive capture as an information problem. But it does not address participant consent, access boundaries, sensitive data, retention, or poisoning [00:08:05]-[00:09:40].
- YC regenerated a living user manual from recorded advice: Blomfield says YC condensed about 2,000 hours of office hours into a 150-page manual. He says it can incorporate new advice each month and become agent context. “Dramatically better” is his assessment. The recording gives no accuracy, contradiction, provenance, or user-outcome evaluation [00:09:40]-[00:11:19].
- Preserve context more carefully than generated software: Blomfield treats business knowledge, data, and skills as durable assets while internal dashboards and workflows become disposable artifacts that can be regenerated as models improve [00:11:19]-[00:12:18].
- Keep people at the high-stakes edge: Blomfield places humans in novel, ethical, emotional, and relationship-heavy situations where agents cannot yet act well. He says he is not sure any company has realized the proposed model across every function [00:08:05]-[00:08:21], [00:12:18]-[00:13:29].