How To Build Superintelligence Inside Your Company

Episode Summary

Executive Summary: The episode argues that “super intelligence” inside a company comes from building AI as an organizational layer, not a co-pilot feature: centralizing context, exposing internal tools to agents, and capturing artifacts like transcripts so knowledge can be reused and improved. YC’s internal agent stack—tool registry, shared database access, skills, and open transcripts—shows how AI can raise productivity, accelerate onboarding, and shift control from developers to users.

Main Topics: Building AI as an organizational layer (Priority: 5/5): Pete and Gary argue that AI should be embedded into core workflows rather than added as a helper. The goal is a shared organizational brain that lets employees tap into collective knowledge and improve how work gets done. YC’s internal agent infrastructure (Priority: 5/5): They describe how YC began with finance workflows and evolved into a broader harness for internal agents, including SQL access, model-file access, and a growing tool registry now used across teams. Context centralization and denormalization (Priority: 5/5): A major theme is that agents work best when important context lives in one place. YC’s single Postgres database and broader internal knowledge systems are portrayed as the equivalent of a ‘big table’ optimized for agent retrieval. Skills, resolvers, and self-improving loops (Priority: 4/5): The discussion compares tool registries, skills, and resolvers across systems like OpenClaw, Claude Code, and YC. They emphasize meta-skills, automatic skill generation, and daily improvement loops using transcripts. Transparency, trust, and organizational culture (Priority: 4/5): The speakers argue that agentic organizations require trust-by-default, egalitarian access, and broad visibility into agent conversations, which both teaches employees and creates social control around safe usage. AI-native software and the ‘horseless carriage’ critique (Priority: 4/5): Pete’s essay is used to critique AI features bolted onto old software. He argues the future is software wrapped around AI/agents, shifting control from developer-defined workflows to user-directed systems. Personal AI vs centralized AI (Priority: 5/5): The episode contrasts open, user-controlled AI systems with a future where a few companies centrally control prompts, models, and access. The speakers frame personal, customizable AI as the preferable path.

Key Arguments: AI should be treated as the building layer for work, not just a copilot, because that lets organizations capture and reuse knowledge across everyone’s tasks. Centralizing internal context—especially in one database or knowledge layer—makes agents dramatically more useful because they can answer arbitrary questions without human bottlenecks. Exposing agents to shared tools and a broad tool registry turns AI from a novelty into operational infrastructure for real company workflows. Recording and reusing artifacts like meeting transcripts allows an organization to convert individual expertise into reusable skills that improve over time. Open access to agent conversations can accelerate learning, create norms, and make AI usage more effective in high-trust environments. Skills should be small, reusable, and resolver-driven; redundant prompts and tools should be collapsed into cleaner abstractions. AI-native software will increasingly look like agents controlling deterministic tools, not deterministic software with a small AI feature layered on top. The best organizational gains come from increasing the floor for everyone—especially onboarding and ramp time—rather than only boosting top performers. A user-controlled AI ecosystem is preferable to centralized corporate control because it preserves experimentation, personalization, and individual empowerment.

Data Points: Time since project started: About 1 year - Pete says YC’s internal agent infrastructure began roughly a year ago and then snowballed into a broader layer. Initial tool count: About 20 tools - The first internal agent system at YC launched with a small shared tool registry. Current tool count: More than 350 tools - Pete notes the registry has grown as every team added tools for its own workflows. Suggested organizational spending: $10,000 to $100,000 per year on tokens - Gary argues organizations need to be willing to invest materially in AI usage to get the benefits. Possible future cost reduction: From $100,000/year to $10,000/year, then a few hundred dollars - Gary predicts token costs will fall quickly, making current spend look temporary. Startup School dates: July 25th and 26th - A promotional insert mentions YC Startup School returning on these dates. Public AI adoption timeline: 2023-2026 - The conversation contrasts the early ChatGPT era with the more mature, agentic era and references 2026 as the present in the dialogue. Time to build earlier project: Half a million lines of code over January to March - Gary describes building a Rails-based system before moving to a more agentic approach. Recent rewrite size: About 40,000 lines of code in three days - Gary contrasts the speed of building with agentic tools versus older manual workflows. Compressed rewrite estimate: Half a million lines of Rails code equals roughly 10,000 lines of TypeScript and 2,000 lines of markdown - Gary uses this to illustrate how much more dynamic the agentic version is.

Pivotal Quotes: "How do you build super intelligence inside a company?" — Gary: Opening framing for the episode’s thesis on organizational AI. "It’s like a shared organizational brain." — Gary: Describing the effect of capturing artifacts and shared context across the company. "By default, the agent conversation is actually globally viewable by any full-time employee at YC." — Pete: Explaining YC’s transparency-first approach to internal agent usage.

Implications: The episode suggests companies should redesign around AI-native workflows, shared context, and reusable skills. The winners will be organizations that empower all employees with open, customizable agents—not just those that bolt AI onto legacy systems.