I suddenly realized my approach to programming has changed. It is no longer mostly “think → write code → debug.” Now it feels more like a continuous dialogue: - conversing with the agent to explore ideas - conversing with other developers to refine judgment - conversing with the coding agent to turn the shared understanding into code - conversing with ... The core skill is shifting from typing the implementation to shaping the conversation.

pie is moving fast: tui / skill / web ui / agents hub (notification mcp) / trigger / cron job / goal mode / ... pie is designed for maximum flexibility and to serve as the foundation for agent applications.

Lovable chose Supabase. Kimi chose TiDB. Both are AI companies, but they are not building the same kind of company. One helps humans generate apps. The other is building an AI system that needs scale. That is why the database choice matters. It shapes what kind of company you can become in the AI era. A great article from @llama_venture

AI agents change the database business model. For years, SaaS apps were built around “one app, one database”. But agents are different. The emerging pattern is: one agent, one database. Why? Because each agent needs its own durable memory, context, tools, permissions, audit trail, and billing boundary. A shared database sounds cheaper at first, until you hit the real problems: - compliance gets messy - tenant isolation becomes fragile - per-agent billing is hard - debugging is painful - noisy neighbors appear everywhere - deleting one agent’s data cleanly is harder than it should be The clean abstraction is simple: Give every agent its own database. The hard part is cost. Can you imagine buying 1 million databases from Supabase, Aurora, or hosting 1 million MySQL/Postgres instances yourself? That model breaks immediately. This is where TiDB Cloud proves something important: “one agent, one database” is not just an architecture pattern. It is a new database business. With TiDB Cloud, we can make a database cheap enough to be provisioned per agent — around $2 or even cheaper per database per year. That changes the unit economics. Agents won’t share one giant memory forever. They will each get their own database. And the winning database platform will be the one that makes millions of tiny, isolated, compliant, billable databases feel boringly easy.

I’ve learned that customers often react better when the story is built in front of them, not presented at them. That’s why I usually prefer a whiteboard over slides. When I draw the architecture live, the thinking becomes visible: the tradeoffs, the flow, the “why” behind the design. The only downside is that my handwriting looks like a failed distributed system — eventually consistent, but not immediately readable. So customers may need to spend a little extra time decoding what I draw, but that process also creates better discussion. They ask questions, challenge assumptions, and co-create the solution with me. In the end, that usually has much more impact than a polished slide deck.

shipped drive9 as a Built-in Connector, right alongside GitHub, Gmail, and Notion. What this unlocks for VM0 users: Every VM0 chat is its own clean sandbox — perfect for safe agent runs. But the work an agent does in one chat (files, images, notes, reports) used to disappear with the chat. Now, with drive9 enabled in your Built-in Connectors, your agent can save that work to a persistent workspace — and any future VM0 chat (or any other agent, on any other runtime) can find it, grep it, build on it. One sandbox per chat. One workspace across all of them. Honored to be in Built-in stack.

AI agents need more than a place to run. They need a safe place to run, and a reliable place to remember. That’s why we built drive9-for-firecracker. Firecracker gives each agent lightweight microVM isolation. Drive9 gives it persistent filesystem state across runs. Clean runtime. Durable memory. One command. Repo: