Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
How to Build a Self-Improving Company with AI
TL;DR
AI should replace the Roman-legion org chart, not just speed it up — The speaker argues most companies still route information like ancient hierarchies, while AI lets you rebuild the company itself as a network of recursive, self-improving loops instead of giving employees “copilots” for a 20–30% boost.
The real asset is company know-how made legible to models — Domain knowledge buried in Slack, email, Notion, office hours, and hallway conversations becomes powerful only once it’s recorded, synthesized, and turned into context, skills, and artifacts an AI can actually use.
YC already has a live example: an agent that improves itself overnight — What started as a database-query sidekick evolved into a monitored system that watches failed employee queries, decides what’s missing, writes code, opens a merge request, gets reviewed by another agent, and deploys fixes before the next morning.
The operating model shifts from 'burn headcount' to 'burn tokens' — The claim is that startups are reaching Demo Day with roughly 5x more revenue per employee than 18 months ago, and soon the bottleneck will be token usage rather than how many people you can hire.
Middle management gets squeezed out; DRIs and builders matter more — The speaker says coordination should increasingly be handled by AI, leaving two key human roles: individual contributors who build and operate, and directly responsible individuals who own outcomes without committees.
Software becomes disposable, but data and context become sacred — Using tools like Codex 55 to one-shot internal dashboards, the advice is to keep every email, recording, and operational breadcrumb forever, while treating the software layer as ephemeral and easy to regenerate as models improve.
The Breakdown
From Roman Legions to AI-Native Companies
The talk opens with a vivid analogy: most companies are still organized like Roman legions, with humans passing orders down and information back up a hierarchy. Borrowing from Jack Dorsey, the speaker says AI breaks that assumption entirely — instead of adding a better engine to the old machine, you can redesign the machine.
Why “AI productivity” is too small a frame
He pushes back on the common 2024 story of AI as a copilot that makes engineers 20% more productive. The bigger idea is extracting a company’s domain knowledge — the stuff trapped in people’s heads, Slack threads, emails, and Notion docs — and making it legible enough that AI can act on it.
The self-improving loop: sensors, policy, tools, quality gates, learning
The core model is a recursive AI loop: sensor inputs like support tickets or telemetry, a policy layer for permissions and logging, deterministic tools like database queries, quality gates like evals or human review, and a learning mechanism that feeds failures back into the system. If that loop can run with minimal human intervention, the company improves “while you’re sleeping.”
The YC moment that made it click
The speaker’s own “holy shit” example came from YC’s internal agent. It began as a simple database assistant for questions like when someone last had office hours with a company, but the breakthrough was adding a monitoring agent that watches every failed query, figures out what tool or index is missing, writes the code, submits a merge request, gets it reviewed, and deploys it overnight.
Product and support loops that can run themselves
From there, he sketches broader use cases: an agent that scans product analytics, finds friction in the funnel, researches best practices, launches an A/B test, picks a winner, and repeats. Or a customer-feedback loop where AI triages suggestions with something like a virtual chief product officer and CTO, deciding what fits the roadmap and shipping changes without waiting on humans.
Burn tokens, not headcount
This shift changes company economics. The speaker says YC is already seeing startups hit Demo Day with about 5x more revenue per employee than 18 months ago, so the limiting factor may soon be token usage, not staffing — and while measuring token usage is “dumb and gameable,” he still thinks it’s directionally useful for spotting who is actually experimenting.
Record everything, because unrecorded work didn’t happen to the AI
His blunt prescription is to make the whole company legible: record partner emails, Slack messages, DMs, office hours, even in-person conversations if possible with phones, clips, smart glasses, or room mics. The line that sticks is basically: if it wasn’t recorded, it didn’t happen to your intelligence.
YC’s self-updating user manual and the case for disposable software
Using roughly 2,000 hours of recorded office hours from the last three months, Haj regenerated the YC user manual over a weekend into a 150-page version the speaker says is dramatically better than the old one written 5–10 years ago. That becomes the pattern: preserve data and business context forever, treat dashboards and internal apps as disposable, and let humans stay at the edges — in emotional, novel, ethical, high-stakes moments like cofounder breakups or sales conversations — where reality still demands a person in the room.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
