How to Build a Self-Improving Company with AI

YC General Partner Tom Blomfield proposes organizing company work around AI loops that learn from their results. Instead of only helping an employee answer a question, an agent would identify why an answer failed and change the tools or context so the next attempt succeeds. He describes YC examples, while acknowledging that he is unsure whether any company has realized this model across every function.

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 00:00–02:24(opens in a new tab).
• Define what each loop observes, can do, and must check: Signals such as support tickets and product telemetry start the loop. Policies specify allowed actions, logging, and when to ask a person. Tools perform actions; evaluations, safety filters, or human review check the result. Failures then become inputs for improving the system 02:24–04:12(opens in a new tab).
• 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 04:12–05:50(opens in a new tab).
• 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 05:50–06:29(opens in a new tab).
• “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 06:29–07:23(opens in a new tab).
• Flatter organizations are his prediction: Blomfield expects AI to replace middle management’s information-routing role, leaving individual contributors and a named person responsible for each outcome. He does not establish that this replaces management's other responsibilities 07:23–08:05(opens in a new tab).
• Recorded knowledge supplies the context: Blomfield recommends capturing emails, Slack messages, conversations, and office hours so agents can use information that would otherwise remain in people's heads. His proposal leaves consent, access control, sensitive-data handling, retention, and maliciously altered inputs unresolved 08:05–09:40(opens in a new tab).
• Compress recordings into useful, updateable guidance: Raw recordings are too large to put into one context window. Blomfield says YC condensed about 2,000 hours of office hours into a 150-page user manual, organized by topic. New advice can be incorporated or discarded each month, and the manual can become agent context. He calls it “dramatically better,” without an accuracy or user-outcome evaluation 09:40–11:19(opens in a new tab).
• 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 11:19–12:18(opens in a new tab).
• 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 08:05–08:21(opens in a new tab), 12:18–13:29(opens in a new tab).

Watch the full talk on YouTube(opens in a new tab).