Nikita Shamgunov on LinkedIn: Here’s the story on how we accidentally created the perfect DB for agents.… (original) (raw)
Here’s the story on how we accidentally created the perfect DB for agents. If you're building Agent AI devtools, get in touch, would love to connect. The Replit Agent launched. It goes beyond code and provisions infra, including Replit Postgres databases backed by Neon. The scale is massive. New databases every few seconds. You need to see it in action. The app it creates is instantly live and operational. Replit did all the work here, we just ship the Postgres. But for us from the database side, it's exciting seeing how perfect a few Neon features are for Agents. 𝐅𝐢𝐫𝐬𝐭: 𝐎𝐧𝐞-𝐬𝐞𝐜𝐨𝐧𝐝 𝐩𝐫𝐨𝐯𝐢𝐬𝐢𝐨𝐧 𝐭𝐢𝐦𝐞𝐬 - When you spend hours writing code, five-minute deploys are no big deal. - When an agent generates the same code in seconds, a five-minute deploy sucks. One thing we got from separation of storage and compute is 1 sec database creates. For developers, this was nice. For agent workflows, it’s a hard requirement. 𝐒𝐞𝐜𝐨𝐧𝐝: 𝐒𝐜𝐚𝐥𝐞 𝐭𝐨 𝐳𝐞𝐫𝐨 Imagine creating a new RDS instance every few seconds. You’d be broke at the first invoice. With Neon, databases scale to zero based on inactivity and wake instantly. Many DBs created by agents might never get queried again after 5 minutes—on Neon, that’s OK. 𝐓𝐡𝐢𝐫𝐝: 𝐀𝐏𝐈 One of our early customers, Retool runs hundreds of thousands of DBs on Neon. Pretty much immediately they needed new API endpoints to manage them: track and limit usage, control config, etc.. Sometimes we wondered: Are we building features for one customer? Turns out agents need the same things, in fact the simpler the API, the fewer tokens needed to control it. SOTA is moving fast, hard to say how Agents will interact with infra 6 months from now, but we’re on board. 𝐈𝐟 𝐲𝐨𝐮’𝐫𝐞 𝐛𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐀𝐠𝐞𝐧𝐭 𝐀𝐈 𝐭𝐡𝐚𝐭 𝐢𝐧𝐭𝐞𝐫𝐚𝐜𝐭𝐬 𝐰𝐢𝐭𝐡 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐃𝐌 𝐦𝐞 𝐈 𝐰𝐨𝐮𝐥𝐝 𝐥𝐨𝐯𝐞 𝐭𝐨 𝐭𝐚𝐥𝐤.
``