How To Build Superintelligence Inside Your Company

Why does it matter? Because YC's nightly agent already beat every partner — and any org can replicate the loop today.

Every night, an agent at YC reads through every employee's agent conversations, finds where it fell short, and rewrites the skills those agents run on. One of those skills — writing the tight two-sentence company description that YC partners have spent years perfecting — now outperforms any individual partner. Pete Kuman, who built YC's entire agent infrastructure, calls this superintelligence inside an organization: not a product, but a composable self-improving loop applied to everything an org does.

• An agent that improves its own skills nightly from real conversation transcripts is already running at YC — and it surpassed individual partners at a core task. • Every organizational workflow can be encoded, run, and auto-improved from its own transcripts — the mechanism is available to any company right now. • The next 18–24 months will determine whether AI centralizes into five corporations or becomes a truly personal computing layer. • The multiplayer agent harness — agent superpowers at the team level, not just for individual developers — is the biggest unsolved problem in AI infrastructure today.

Superintelligence inside a company is a self-improving skill loop, not a product you buy

A skill that now outperforms any individual YC partner didn't emerge from fine-tuning a model. It emerged from a feedback loop: a partner encoded how to distill a company into two crisp sentences. Teams used it. Transcripts accumulated. A nightly agent read those transcripts and meta-prompted itself: given what you learned from this context, improve the skill. It did — noticeably.

"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." The dream cycle is already live.

"How do you build superintelligence inside a company? You do that on everything you do. And it's not more complicated than that."

The two-sentence description is one pinprick in the fabric of what any organization does. But the mechanism scales uniformly. Every writing format, every finance workflow, every onboarding script is a candidate for the same loop: encode the skill, run it, collect the transcripts, meta-prompt nightly. Stop treating AI as a tool you reach for. Treat every workflow as a skill that improves itself from its own usage.

We are at the Apple I moment — and the next 24 months will determine whether AI is a personal computer or a new mainframe

In 2034, there's a coherent version of the future where five corporations hold the AI layer of your entire computing existence and you can't change a single prompt. "Is 2034 going to be like 1984? The 1984 case would be we have centralized control... they don't let you run your own prompts. Like they literally do the Gmail thing, but for your whole computing existence." A small priesthood with institutional control of intelligence — just as the 1960s mainframe priesthood controlled computation before personal computers broke it open.

The alternative is the homebrew computer club. Wozniak soldering Apple I boards in a garage. Five hundred units sold to enthusiasts who didn't yet have a word for what they were building. "I think we're at the Apple I moment right now. We are coming up with the primitives."

ChatGPT reaches a billion users with tightly controlled MCP access. The tools that offer real control — self-hostable, model-switchable, prompt-editable — remain kits for enthusiasts. The window is 18 to 24 months. Defaulting to managed AI products today is a quiet vote for centralization. "There's this other alternative where I want a billion people to actually control and program for themselves. This should be an extension of yourself." The people watching will build the infrastructure that decides which future lands.

The single-player agent era is ending — nobody has built the team multiplayer harness yet

Claude Code, Codex, Pi, OpenClaw, Hermes — every agent harness that made individual developers feel superhuman was designed for one human on one machine. That was the right call: start simple, prove the value. But it means almost all the organizational leverage from agents hasn't been built yet.

"It feels like we're still kind of in the single player era of agents. One of the big problems that I don't think has been solved well yet by anybody is the multiplayer harness. It's enabling that kind of superpower, but on a team or an organizational level."

A single developer with a coding agent can compress days of work into hours. Now picture a team sharing the same tool registry, the same context, the same skill library — where every improvement one person makes propagates immediately to everyone else. That's the organizational flywheel. It exists in theory. It hasn't been productized.

What YC has built — shared registry, common database, broadcast conversations — are early multiplayer primitives. The harness layer that makes an entire org feel what a solo developer feels today hasn't shipped from anyone. That gap is the highest-leverage infrastructure bet available right now.

One Postgres database beat twenty SaaS tools — because it collapsed the cost of asking hard questions to zero

Collapsing the cost of a complex query from hours to seconds didn't just make existing work faster — it changed which questions teams dared to ask at all.

YC runs on software it built itself, everything sitting in one Postgres database: every company funded, every founder, every financial transaction, every CRM note. When an agent got read-only SQL access, the effect was immediate. "It dramatically increased the number of questions that we would ask and dramatically increased the scale and complexity of the questions that we would dare to ask."

Queries that previously required hours of SQL work — which investors backed space companies in the last four batches? — dropped to seconds. Before, teams simply didn't bother unless the question was critical. After, volume exploded and the questions grew more ambitious. Jevons paradox applied to organizational intelligence: make something cheap enough and usage doesn't just increase, it changes shape entirely.

"Just like a coding agent inside a monorepo just tends to be much more efficient, watching our agents operating on our single database that has everything in one schema" — the coherence compounds. The lesson isn't build your own software. It's that fragmented context across twenty SaaS tools is the primary ceiling on organizational AI. Audit where your critical context lives before investing in agent features.

A shared tool registry — not a better model — is what turned generic AI into something useful at work, and it compounds every month

Twenty tools at launch. More than 350 today. Every team at YC adding their own: finance books journal entries, events get coordinated, partners manage office hours. Every addition lands in the shared registry — available in internal agent harnesses, in Claude Code on individual laptops, in whatever harness someone builds next.

"The tool registry is where most of the YC-specific stuff lives. The tool registry is what turns these agents into something that's useful at work."

The governing principle: DRY and MECE. If two tools do overlapping work, collapse them into one with parameters. If a gap exists, fill it. A "Check Resolvable" meta-skill continuously validates the registry stays clean. Kuman built this pattern at YC; Friedman arrived at the same structure independently in OpenClaw. Same primitive, converged on in parallel — which suggests it's structurally right, not just locally useful.

The flywheel is structural: each new tool makes every agent in the organization smarter. That means the registry is first-class engineering infrastructure, not a side project. It needs ownership, maintenance, and contribution as a standard part of every team's workflow — not as an afterthought.

Most enterprise AI is a horseless carriage — old software shapes with a language model bolted inside instead of an agent driving

Most companies building "AI features" are slotting a language model into an old workflow — prompts hidden from users, path predetermined by developers, control locked away. "The potential for AI is to shift control of software from the developer to the user." Most products are doing the reverse.

The Gmail AI writer is the canonical example: a button appears, text generates, the user can't see the prompt or adjust the persona. A feature, not a shift in power. The gap between that experience and Claude Code on your own machine — where you can do anything — keeps widening as the two trajectories diverge.

The inversion AI-native software requires: put the agent in front, wrap deterministic tools underneath as its levers. The agent determines the path; the tools provide the guardrails. "The best AI software that I've used tend to be very small and just add the smallest amount of code ahead of time that you need in order to let the model shine."

When users can modify prompts and the agent has genuine agency, the developer doesn't need to anticipate every use case in advance. Product surface area shrinks; user power grows. If the workflow is predetermined and the prompts are hidden, you haven't shifted control — you've just upgraded the carriage engine.

Broadcasting every employee's agent conversation publicly solved knowledge transfer and security simultaneously — through social pressure, not policy

Agent conversations at YC are globally visible to every full-time employee by default — broadcast to a Slack channel anyone can join. The security instinct screams against this. The results argue otherwise.

"People learned how to use it from watching how other people use it." When a partner used the agent in a way no one had seen before, colleagues watched and replicated. Adoption spread laterally, faster than any training program would have achieved.

The security mechanism turned out to be social, not technical. "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. That, as we learned, has been reasonably effective at keeping private information private." When your agent conversation is visible to every colleague, you don't use it to do something you'd be embarrassed to have them see. Social accountability outran policy.

The standard instinct — restrict AI access to protect sensitive information — produced less protection and less adoption than radical transparency did. The prerequisite doing heavy lifting in this sentence, as always, is a genuinely high-trust environment.

Egalitarianism and trust-by-default aren't cultural aspirations — they're technical prerequisites most organizations can't retrofit

The infrastructure for a self-improving organizational AI is available today. The cultural infrastructure is not — and capital alone can't close that gap.

"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 neither of those things actually are most organizations in the world."

By default, a company is command and control. Leadership gets the powerful tools; line-level staff get the guardrailed version. Prompts are hidden. Agent conversations are siloed and audited. Every one of these defaults works against the self-improving loop. You can't have a tool registry every team contributes to if only certain teams are trusted to touch infrastructure. You can't run a nightly dream cycle on conversation transcripts if those conversations are locked.

Incumbents with resources are already falling behind — not because they lack models or compute, but because retrofitting egalitarianism into a command-and-control org is structurally hard in a way money can't fix. "By default, a company is command and control. By default, maybe the leadership gets it... the line level people, the staff people don't."

"The people who are watching are going to be the people who build all these things in society. So we better choose well." Encode egalitarian AI access and trust-by-default before you scale. The defaults you ship with are the ones you'll live with.

The compounding gap is already opening — and by the time it's obvious, it'll be too late to close

Token costs running $10,000–$100,000 a year today will fall to a few hundred dollars within two years. Model costs decline. Compute costs decline. What doesn't decline is the institutional muscle memory built by organizations already running nightly self-improvement loops across hundreds of tools. An org that has done this for three years doesn't just have better tools — it has a different organizational nervous system, refined daily. When the costs normalize and every competitor can finally afford to start, the question won't be who spent the money. It'll be who spent the time. Late movers can buy the same models. They cannot buy three years of compounding.

---

Topics: artificial intelligence, organizational AI, agent infrastructure, YC, internal tooling, knowledge management, AI strategy, startup culture, AI native software, personal AI