init() fireside recap: Nikita Shamgunov on getting your database picked by an agent

extract

init() fireside recap: Nikita Shamgunov on getting your database picked by an agent Databricks VP Nikita Shamgunov at WorkOS init() on Neon's one-URL widget, the day Replit Agent created 13,000 databases, and how to get recommended by agents. Replit had twenty to thirty million accounts when Neon finished the integration. It produced about ten databases a day, all of them scaling to zero, and revenue that rounded to zero. Then Replit launched Replit Agent, and Neon created 13,000 databases in a single day. Nikita Shamgunov, VP of Engineering at Databricks, told that story to Michael Grinich in a fireside chat at init() on October 7, 2026, at SFJAZZ Center in San Francisco. The full init() 2026 recap covers the keynote and the rest of the day, and the whole program is on the init() event page. Shamgunov worked on SQL Server at Microsoft and then at Facebook, co-founded SingleStore, and started Neon in 2021. The chat kept circling back to one tension: the discipline that made Neon small enough for an agent to adopt, and the fear of staying that small. A database narrow enough to be a URL Shamgunov's read on SingleStore is that it scaled past a hundred million dollars while carrying a few architectural decisions that were hard to reverse. The company set out to do everything database, operational and analytical workloads both, and built a lot of technology without producing something instantly useful to a specific audience. Neon started from the opposite end. Separating storage and compute meant writing new storage from scratch, which was slow but became the defensibility. It bought two properties, serverless and branching, though the team only knew about one of them going in and found the second later. Shamgunov called Supabase Neon's biggest competitor, said it is ahead of Neon, and said it still lacks proper branching and proper serverless because that architecture is hard to build. Then came the subtractions. Neon counter-positioned against AWS Aurora, which supported MySQL and Postgres, by supporting Postgres only, and went cloud-only despite shipping an open-source repo. About nine months in, the team narrowed again to serverless only: And then we were like, database is just the URL, it's a useful URL because you can build an app on it. And so that's the widget, right? Database is just the URL, is the widget. The reception matched the focus. The open-source repo collected stars immediately, the waitlist swelled daily, and the site reached number one on Hacker News after leaking there with no promotion. Shamgunov contrasted that with flying to New York in a suit to pitch Goldman Sachs and Morgan Stanley: at Neon it was a bunch of nerds twisting knobs on the backend. The widget becomes somebody else's feature The catch arrives on a delay: Of course, then the thinking is, now if you're a widget, and you're a widget for a very long time, then you will become somebody else's feature. His example was PagerDuty, which at one point had half a billion in revenue, was profitable, and was trading at half a billion, and is now at a billion. Not staying a widget means going multi-product, and he quoted ServiceNow CEO Bill McDermott's version of it: you got to build the whole platform. What AI changes, he thinks, is the clock. The gap between a useful widget and a platform keeps shrinking. Friends, foes and shelf space Being narrow made partnership strategy nearly mechanical. Shamgunov divided the world into friends, foes and don't-care. Any database company was a foe, including Supabase, PlanetScale and AWS Aurora. Anyone with a platform and no database was a friend: Vercel, Retool, Replit, Lovable, Railway, Render, plus frameworks, ORMs and libraries. He described the result as shelf space that developers walk past, with Neon's job being to sit on every shelf it could reach. The Replit integration took about a week, helped by the APIs Neon had already built for Hasura. Modern enterprise sales to the tech industry, he said, is a joint Slack channel where both sides wire things together. That is how the integration got built, and then it did nothing. Investors told Neon to focus on Vercel instead, and Vercel did deliver far more traffic. The Replit Agent launch in fall 2024 reversed it. Creation tapered after the 13,000-database day but never fell back under ten, and Shamgunov now puts it at 50,000 to 70,000 databases per day from Replit, hedging the exact figure but not the order of magnitude. The people creating them are mostly non-technical and have no idea they are provisioning a Neon instance. Three levers decide which database an agent provisions: what you publish, what your partners agree to, and whether you are in the system prompt. Harness relations Which brings up the uncomfortable part. Sometimes Claude recommends Supabase first and sometimes Vercel first, and when it recommends Vercel first, Neon wins every single time. Grinich asked whether that is something you pray to the LLM gods for. Shamgunov said there are three things to do before praying: publish a lot of factual material and make sure your developer experience actually matches it, work your partners, and influence harnesses. So arguably, every SaaS tool is becoming a harness. His definition: a tool where you do work, it makes LLM calls, and it wraps those calls in constraints and agentic loops. That makes a harness another piece of shelf space. So you go to harness owners and ask them to add a prompt putting you ahead of the competition, assuming the competition is making the same call. He called the practice harness relations rather than developer relations, and said it is really partnerships. When an owner declines, the remaining option is influencing the model itself through agent optimization, the AEO successor to SEO and GEO, measured with evals: So you have your evals, you have all your prompts, and you see who's recommending what, and then you're trying to understand why, and then change this. Weights are heavy. Neon is living the lag right now. It launched a Supabase competitor called Postgres+, and models still answer as though Neon only does Postgres until you hint otherwise. A system prompt has no lag, which is why Shamgunov takes one wherever he can get it. The Replit relationship is an agreement that Replit always uses Neon, with the logo in the connection string. The third lever is speed. Something new ships for agents most days, and you have to support it immediately: So you either are on the agentic learning curve or you're not. And if you get off the agentic learning curve, then catching up is very, very hard. His example of something to support was auth.md, from WorkOS. He also admitted Neon was not yet in OpenAI's marketplace, with the wiring underway after a conversation in a shared Slack channel. Keeping up means funding teams to chase new shelf space weekly; at Neon that is a technical DevRel team that ships, publishes and measures itself on adoption. The numbers he gave: MCP and skills adoption growing in the twenty-some percent range week over week, against drivers downloaded about five million times a week, with the former possibly overtaking the latter. Grinich said WorkOS renamed its developer experience team to agent experience for the same reason. The rename follows the same logic as the rest of the list: design for whatever is reading your docs and choosing your product. Who holds the interface Shamgunov's durability argument is a barbell. In the long run the valuable positions are whoever holds the interface to the end user and whoever owns infra, and in a world where code is nearly free, infra means inference. Dev tools used to hold that interface. He named Neon, Vercel, WorkOS and Stripe as direct interfaces to the developer, with Stripe as the bar-none best at clean APIs, docs and design. The problem is what happens when an agent becomes the one reading the docs: But just like in a hotel, the front door is very pretty but the service elevator is not because it doesn't need to be. If the interface moves up into the harness, power moves with it, toward Claude Code, Codex, Cognition and Factory, in a market he says seems unlimited. His consolation is that data is still infrastructure and data is hard, at least at this point in time. His structural answer is a platform that crosses both use cases and users, the way Microsoft sells to developers and finance and HR and IT, and the way Databricks does at north of a 7 billion run rate. Why Neon sold The widget-to-platform transition is risky because you have to find product-market fit for a second act while the first one is still growing. Neon had the obvious adjacencies mapped: auth, observability, product analytics, data warehousing, hosting, the backend-as-a-service surface Supabase covers, and sandboxes. Then Databricks called. Neon was not shopping, Reynold called on a Friday night, and they shook hands by Wednesday, with a term sheet and a Databricks partnership both in flight at the time. The pitch was that Databricks already had the platform Neon would otherwise spend years building, plus enterprise distribution beyond self-serve. Shamgunov says revenue was small at acquisition and is more than 10x now, a number he flagged as a sandbag, and that the price would probably have been much higher had they sat tight. On-ramps, not a funnel He is candid that the PLG revenue may not be the durable part, because so much of it comes from agentic app building. He and Michele from Replit have discussed over dinner whether the industry is hitting peak apps, with mini apps getting absorbed into ChatGPT, Claude and the AI factories. Neon still has more PLG revenue than enterprise revenue, barely, and he expects enterprise to exceed PLG by 10x eventually in a hundred-billion-dollar database market. That forecast cuts both ways, since the base it is measured against is the revenue he just said might not hold. The mental model he uses for getting there is on-ramps rather than a funnel. Crossing a capability threshold opens an on-ramp; putting traffic on it is a separate problem you solve afterward. The enterprise on-ramps are the familiar list: high-end performance, compliance, GovCloud, FedRAMP, security, bring your own keys, store data in your own bucket. The destination is being default Postgres, which he was careful to call a vision and not today's reality. Two customer stories showed what the on-ramps feel like from the other side. Meta is a multi-million dollar user that arrived the same way a hobbyist does, through an easy Vercel integration, spread to other team members inside the product, and only later asked for compliance features. A large real estate firm built all its internal tools on Neon, then discovered it was not allowed to pay by credit card; told that the non-credit-card path starts at 50k, it agreed on the spot. Shamgunov's rule of thumb from SingleStore was that crossing 700k triggered higher-level approval ten years ago, and multi-million deals pull in VPs and CFOs. Do you have the right to win Asked for investor-hat advice (he spent time at Khosla Ventures partly to avoid becoming a database guy forever, and says his investments cleared a billion in returns to the firm), he gave the barbell again: be where the work happens, or be on infrastructure you believe will last. Building is cheaper now, so everything will be more competitive. Once the market decides a category exists, he says, you have to ask honestly whether you have the right to win it: I think once the market decided that this is a category, it's just so much easier for the end user to think about in these terms. Or the agent, I guess. AI factories look like a category to him, and inference platforms like Baseten and Fireworks look like another. His bitter lesson was telling himself HTAP was a category when the real one was data and analytics in the cloud, which Snowflake and Databricks ran away with. The signal Neon's evals-over-prompts loop is the piece worth copying directly. It tracks who is recommending what and then works out why, and it has produced evidence that the recommendation decides the sale: when Claude names Vercel first, Neon wins every time. If an agent can provision your product, that recommendation path is a funnel you can instrument. Postgres+ is the cautionary half of it, because you can ship an entire new surface area and still get described as the old product until someone hints otherwise. Narrowness got Neon onto the shelf and into the system prompt. It is also what put Shamgunov on a call on a Friday night and a handshake by Wednesday.