crosstalktalks, distilled into a living wiki

How to Build a Self-Improving Company with AI (Tom Blomfield, YC)

Y Combinator · watch the talk ↗ · updated Jun 10, 2026

Tom Blomfield, Y Combinator General Partner, argues that the copilot / productivity-boost mental model for AI is obsolete and that founders should instead redesign their entire company as a set of recursive self-improving AI loops — layered feedback cycles that detect failure, generate fixes, and deploy them without human intervention. Drawing on live YC internal examples, he outlines the architectural principles needed to reach a state where the company improves while its founders sleep.

Key Takeaways

  • The hierarchical organization (Roman-legion model) exists to route information through humans; AI eliminates the need for that structure. 0:00
  • AI copilots as productivity enhancers (+20–30%) are "last year's model" — the real leverage is autonomous self-improvement loops. 4:26
  • YC portfolio companies are already reaching demo day with ~5× more revenue per employee than 18 months ago. 6:34
  • Middle management is over; only individual contributors (ICs) and directly responsible individuals (DRIs) remain necessary. 7:23
  • Software is ephemeral; domain knowledge / business context is the durable, valuable asset. 11:54
  • Humans still matter for novel situations, ethical calls, high-stakes emotional moments, and physical-world interfaces (e.g., sales conversations). 12:39

The Broken Mental Model: Copilots and Roman Legions

Blomfield opens by crediting a Jack Dorsey tweet for crystallizing the problem: today's companies are organized like Roman legions — nested hierarchies where humans are the conduit for information flowing up and down 0:19. The dominant AI response has been to bolt a "more powerful engine" onto this old structure through copilots that make engineers 20% more productive 1:11. Blomfield explicitly calls this a broken way of thinking about AI 1:22 — it preserves the hierarchy rather than replacing it.

The Recursive Self-Improving Loop Architecture

Blomfield proposes replacing the hierarchy with feedback loops composed of five layers 2:46:

  1. Sensor layer — raw inputs: customer emails, support tickets, cancellations, product telemetry 2:50
  2. Policy layer — decision rules: what the agent can act on autonomously, what requires human sign-off, what must be logged 3:07
  3. Tool layer — deterministic APIs (database queries, calendar lookups) the agent can call 3:15
  4. Quality gate — evals, safety filters, human review for high-risk actions 3:30
  5. Learning mechanism — failures feed back to the top, driving the next iteration 3:37

When every step runs with minimal human intervention, the system self-improves continuously 3:44.

Concrete YC example: A query agent was given tools to look up office-hours history. A monitoring agent layered on top watched every employee query, identified failures overnight, wrote new code, opened a merge request, had another agent review and merge it, and deployed — so the same query succeeded the next morning without any human involvement 4:29–5:07. Blomfield calls this his personal "holy shit moment" 5:07. He extends the pattern to a self-optimizing product loop (analytics → friction detection → A/B test → deploy → repeat) 5:34 and a customer-support loop (triage agent acting as virtual CPO/CTO, shipping feature suggestions overnight) 5:54.

Organizational Implications: Tokens, Not Headcount

  • Burn tokens, not headcount 6:29: token spend is the new constraint, not hiring. Blomfield recommends tracking token usage across the org directionally (while acknowledging it becomes gameable as a formal KPI) 6:49.
  • Middle management is done 7:23: AI handles coordination; the only roles that matter are ICs/builders and named DRIs. Blomfield explicitly disagrees with one of Dorsey's three roles and drops it. 7:29
  • Founders at early-stage companies have no excuse not to build AI-native from day one 13:16.

Making the Organization Legible to AI

Blomfield's operational prescription is total capture: "If it is recorded, it happened to the AI. If it did not get recorded, it did not happen to your intelligence." 8:39

YC now logs all partner emails into a central database, records every Slack DM, and has been recording all office hours for ~3–4 months 8:25. Because 100,000 hours of audio cannot fit in a context window, the pipeline requires diarization, aggregation, and synthesis into retrievable breadcrumbs 9:20.

YC user manual regeneration example: YC partner Harj Taggar used 2,000 hours of recorded office hours to regenerate the outdated (5–10-year-old) user manual over a single weekend, producing a 150-page document that Blomfield judges dramatically better 9:47–10:11. The manual now updates monthly: new advice is compared against the existing document and either incorporated or discarded, making it a living company brain 10:13.

Software Is Ephemeral; Context Is the Moat

Blomfield argues that internal software (dashboards, workflows) should be treated as entirely disposable — generated on demand via tools like Codex, thrown away when models improve, and regenerated from the same instructions 11:00–11:53. The permanent, precious asset is structured data and domain know-how: "The business context and skills are the valuable part. The software on top of it is ephemeral." 11:57

Human roles are repositioned to the edge of the company brain — interfacing with reality in places models cannot yet reach: in-person conferences, novel situations, ethical dilemmas, high-emotion co-founder crises, and face-to-face sales conversations (which Blomfield predicts will remain human for the next 20 years) 12:27–13:01.