How To Build Superintelligence Inside

Podcast: Lightcone | Participants: Garry Tan, Pete Koomen

1. Key Themes

The "Multiplayer Agent" Problem Is the Next Frontier

While single-player agent harnesses (Claude Code, Codex, OpenClaw) have matured rapidly, the organizational-level equivalent remains unsolved. YC has been quietly pioneering this by building shared tool registries, common context layers, and broadcast agent conversations — essentially creating the infrastructure for collective AI intelligence.

"It feels like we're still kind of in the single player era of agents where the harnesses that have gotten really popular, Claude Code, Codex, Pi, OpenClaw, Hermes, they're all designed to be used by a single human running on a single machine... one of the big problems that I don't think has been solved well yet by anybody is the multiplayer harness." — Pete Koomen 00:12:12

Organizational Superintelligence Is a Compounding Loop, Not a Feature

The mechanism for building superintelligence inside an organization is concrete and repeatable: encode a skill in a prompt, use it, collect the transcripts of using it, feed those transcripts back to auto-improve the skill nightly. Repeat for every workflow. The skill becomes better than any individual human practitioner.

"How do you build superintelligence inside a company? You do that on everything you do. And it's not more complicated than that. Like you literally just compose everything that you do and any given thing that any given person can do, you combine that in aggregate and in this particular process. And like you have a super organization." — Garry Tan 00:24:18

"We have this general agent that every night will go and read through all of the agent conversations that employees have had and look for things that could have done better and pieces of context that if it had up front, it would have done more efficiently." — Pete Koomen 00:19:26

Centralized Context Is the Unfair Advantage

YC's ability to unlock agentic power came from a single architectural decision made years earlier: running everything on one Postgres database. Organizations fragmented across SaaS tools are structurally disadvantaged. The new version of this is intentional "denormalization" of context into an agent-optimized knowledge layer.

"All of that software sits on one Postgres database that has everything that's important to YC's world in it... when all of that context is in one place, with a little bit of additional information about how the schema is laid out, an agent can go and answer arbitrary questions about our business." — Pete Koomen 00:08:09

"You basically have to take that you're going to denormalize it and you're going to put it in a format that is optimized for agent retrieval and understanding." — Garry Tan 00:11:29

2. Contrarian Perspectives

Radical Openness — Not Security — Is the Right Default for AI Systems

Most organizations lock down AI access in the name of safety. YC found the opposite: making all agent conversations publicly viewable internally solved the security problem through social accountability, while simultaneously accelerating adoption via peer learning. Restricting access destroys value without meaningfully improving safety.

"One of the things we talked about earlier was this trade-off where these agents are at their most powerful when they are given unrestricted access to lots of context, which runs counter to the way most organizations work. It turns out that by defaulting to public broadcast for these conversations, you kind of institute a bit of a social control on what people can do with it." — Pete Koomen 00:28:39

"Perhaps foreshadowing subsequent things like OpenClaw where it turns out that like the thing that was hampering the world was being worried about security and privacy and all the things that could go wrong. And when you like worry a bit less, you're like, oh my God, these things are unbelievably powerful." — Garry Tan 00:06:59

Chat Is the Final Interface, Not a Placeholder

The conventional wisdom in 2023-2024 was that chat was a temporary, primitive UI that would be replaced by something more sophisticated. The contrarian conclusion is that chat is actually optimal because it most closely mirrors human thought expression — and constraining AI into rigid UI boxes actively limits its capability.

"Why chat is probably the better interface is because it's the closest thing to human language. And human language and writing is basically the closest thing to expression of thinking. So chat is the closest stepping stone to clear intelligence." — Garry Tan 00:35:32

The Best AI Software Is the Smallest

Counter to the instinct to build comprehensive, feature-rich AI products, the most powerful AI software is minimal — a thin layer that exposes the model's full capability rather than wrapping it in developer-controlled determinism.

"The best AI software that I've used, whether it's inside of YC or tools that others have built, tend to be very small. And just add kind of the smallest amount of code ahead of time that you need in order to let the model shine." — Pete Koomen 00:38:46

The $100K Annual AI Budget Is a Time Machine, Not an Expense

Spending heavily on tokens now to build agentic infrastructure is not a cost — it's buying yourself 2-3 years of organizational advantage before the technology becomes commoditized and universally adopted.

"Basically what I realize is it allows you to live in 2028... what you spend $100,000 or a million dollars a year on now, it will be commonplace in two years. It won't cost $100,000 in a year, it'll cost $10,000. And the year after that it'll be like a couple hundred bucks... there's a one-time time warp where you can leapfrog every incumbent, all Fortune 500s, all startups that exist by doing this." — Garry Tan 00:29:54

Egalitarianism and Default Trust Are Structural Requirements for AI-Native Orgs

Most people assume AI transformation is a tooling problem. The real blockers are cultural: hierarchical, need-to-know organizations are structurally incompatible with the architecture required to build organizational superintelligence.

"It betrays two traits of truly agentic, like 1000x super intelligent organizations that I would not have necessarily guessed would exist but are now like must exist. If you want to create this type of organization, you have to be relatively egalitarian and you also have to be trust by default. And then neither of those things actually are most organizations in the world." — Garry Tan 00:29:15

3. Companies Identified

OpenClaw An agent harness/framework, likely internal or early-stage, that allows users to run their own agents with full access to their own databases, prompts, and systems. Mentioned repeatedly as a leading example of the "personal AI" paradigm — user-controlled, open, and extensible.

"The ability to run your own software, to change your own prompts, to test all of it, to have your own private repo that like you know is only yours. To be able to choose which model to use... to me that's sort of the white pill for AI." — Garry Tan 00:43:55

Hermes Agent An agentic framework mentioned alongside OpenClaw as a leading multiplayer/personal agent harness. Notably, Hermes already has an automated skill-creation capability ("it makes skills automatically").

"Hermes actually already has Skillify. They call it something. It's like it makes skills automatically." — Garry Tan 00:15:25

Pi (Open Source Harness) A minimal open-source coding agent harness. Described as a beautiful, self-extending piece of software — the agent can use Pi to modify Pi itself. OpenClaw is built on top of it.

"It's this beautiful piece of software, which is just like the smallest possible coding agent. You can use Pi to modify and extend Pi, right? And it's this kind of idea of like self-extending and self-referential software." — Pete Koomen 00:39:55

Perplexity Cited as currently offering the most open consumer AI interface (Perplexity Computer), though still limited compared to OpenClaw/Hermes.

"Perplexity Computer is probably the best version of it. But it's still like, you know, pretty limited compared to what you could do with OpenClaw and Hermes Agent." — Garry Tan 00:43:28

Cursor / Windsurf Cited as the first generation of agentic coding tools that catalyzed the realization of the gap between old software development workflows and new agentic possibilities.

"This was right around when Claude Code was introduced. It felt like this was giving me superpowers... you had kind of the first generation Windsurf and Cursor that were well-established by this point." — Pete Koomen 00:04:04

4. People Identified

Pete Koomen General Partner at YC, creator of Optimizely (pioneering A/B testing platform). Architect of YC's entire internal AI agent infrastructure, including the tool registry (now 350+ tools), the shared agent loop, and the nightly self-improvement dream cycle. Author of the viral essay "Horseless Carriages."

"He has gone on to create all of our agent infrastructure at YC. So literally all of our harnesses and how we use AI internal to YC." — Garry Tan 00:00:38

Andrej Karpathy Mentioned as an originator of key concepts being independently rediscovered: "auto research" (self-improving agent loops) and "knowledge LLM wikis." His ideas are described as foundational to the organizational AI stack.

"Auto research from Karpathy again... Slash goal now in Codex. Like they've incorporated it too." — Garry Tan 00:19:22

Boris (Claude Code) The lead on Claude Code at Anthropic, noted for an exceptional and rare obsession with simplicity and minimalism in AI product design — described as a key differentiator for Claude Code's success.

"Every time Boris comes and speaks at YC... one of the things that really stands out is how obsessed he is with simplicity and with just like making the product as small as possible." — Garry Tan 00:39:41

Jack Dorsey Cited as a visionary example of an operator applying the same organizational AI transformation principles at Block — attempting to build a company-level AGI focused entirely on payments.

"I'm sure you guys have heard Jack Dorsey talk about what he's doing with Block. He basically is trying to turn Block into a mini-AGI around helping people in the world make payments to one another." — Garry Tan 00:23:02

Tom (YC Partner) An unnamed-by-last-name YC partner who wrote the original "two-sentence description" skill — a concrete early example of encoding institutional expertise into a reusable, auto-improving agent skill.

"Tom, one of the partners here, wrote a skill that teaches an agent how to take some context about a company and condense that into a two-sentence description." — Pete Koomen 00:21:50

5. Operating Insights

The Tool Registry as Organizational Operating System

Rather than building purpose-specific AI tools one at a time, build a shared, living tool registry that every team contributes to. Over time it becomes the connective tissue of the entire organization's AI capability — and critically, it works for both internal agents and individual developer harnesses simultaneously.

"The tool registry is where most of the like YC specific stuff lives, right? The tool registry is what turns these agents into something that's useful at work... there are more than 350 today. Every team is adding their own tools... And now once these all exist in one place, you can make them available to these internal agents that we've built. But you can also make them available to Cloud Code, you know, running on our individual machines." — Pete Koomen 00:14:04

Broadcast Agent Conversations to Accelerate Adoption

Make internal agent conversations visible to all employees via a Slack channel or equivalent. This simultaneously trains adoption (people learn by watching power users), creates social accountability that self-regulates misuse, and eliminates the need for heavy access controls.

"Every agent conversation was broadcast internally to a Slack channel and anybody could join that Slack channel and look and learn... I remember when you started using it really heavily, you were super creative with the things you were doing with it. And a lot of us watched that and was like, oh, wow, I didn't even occur to me you can do that now." — Pete Koomen 00:28:10

Use MECE + DRY as the Resolver Optimization Framework

When building a skill or tool registry, systematically audit it against two principles: DRY (don't repeat yourself — no duplicate tools) and MECE (mutually exclusive, collectively exhaustive — full coverage, no overlap). This is the architecture of an optimal resolver that agents can navigate efficiently.

"I run Check Resolvable, which is like, look at all of the other skills and tools that exist. And is it dry? Don't repeat yourself. And is it MECE — Mutually exclusive, collectively exhaustive... it's bad to have 10 skills that do all the same thing. It's good to have one skill or one tool that has parameters." — Garry Tan 00:16:21

6. Overlooked Insights

The "Floor Raising" Effect: AI as Automatic Apprenticeship at Scale

Briefly mentioned almost as an aside, this is actually one of the most significant organizational insights in the episode. AI agent infrastructure that encodes how your best people work doesn't just help star performers — it dramatically compresses onboarding time and eliminates the bottleneck of senior talent being too busy to mentor. The best people's judgment becomes infinitely scalable without their time.

"You could have a new employee joining and maybe would have taken them six months to ramp up. But with this, it's sort of like they automatically get a lot of the contacts from the company working and they know how the best people and the star players in the organization do things. By apprenticeship, automatically with AI instead of because partner time is expensive or sometimes the best people in an org, they're very busy." — Garry Tan 00:31:00

This is non-obvious because most AI deployment conversations focus on productivity of existing staff — not the structural elimination of the mentorship bottleneck that caps organizational scaling. The implication for investors: any company that cracks "AI-native onboarding" as a product category is addressing a massive, underappreciated enterprise pain point.

Meeting Transcripts Are Underutilized Raw Material for Skill Improvement

Mentioned only briefly, the mechanism of feeding meeting transcripts back into skill prompts to improve them is far more powerful than it sounds. It means the latent expertise exchanged in normal organizational conversation — office hours, group critiques, feedback sessions — can be automatically harvested to improve AI systems without any additional human effort.

"A couple of the other partners took a meeting that they had with a group office hours... and just went through and had every founder try their hand at a two-sentence description and kind of gave them feedback and input... handing that back to the agent and saying, given what you've learned by reading through this context, improve the two-sentence description skill. And they got noticeably better after that. Like this thing is now better than I am." — Pete Koomen 00:22:15

The overlooked implication: the wave of meeting recording tools (Otter, Fireflies, Granola, etc.) are not just productivity tools — they are training data pipelines for organizational AI. Companies or investors who understand this reframe the category entirely. The value isn't the transcript; it's the compounding improvement to every downstream skill the transcript feeds.