This is me building @BlockRunAI

Last night I had the pleasure of attending the fireside between @gakonst (Paradigm) and @rauchg (Vercel). Guillermo shared so much valueable insight, specially: MCP events and multi-agent collaboration look promising. 1. The open-model flip. Ten months ago open-weight models were close to 0% of tokens on Vercel AI Gateway. Last month they were about 70% of tokens, with proprietary at 30%. Spend still skews proprietary. The pattern: plan with a frontier model, execute with an open one. 2. Open follows closed in weeks. When a new small, fast proprietary model launched on the gateway, it became the most requested model in its history. A few weeks later there were ten open alternatives. He called this the new normal. 3. "Everything is code." That's why every lab pivoted to coding agents. If the model can code, it can drive Blender, run a sandbox, automate anything. Vercel's answer to lab competition: real open source and developer choice, which labs are disincentivized to offer. Companies end up using 4 or 5 models and need a layer that cuts across them. 4. Agents are the new users. Vercel now designs for agents, markets to agents, and asks agents what they think. Their DX team is now "agentic developer experience" and its goal for the year is verifiability: linters plus skills, correctness loops. They shipped a site that tells you whether agents can read your website. Most can't. 5. Rust is thriving for the same reason. It was less palatable to humans; agents persevere through anything, and they benefit from the correctness loops a strict compiler gives them. Vercel is also exploring what you can add to TypeScript to make it sounder and faster for agent-written code. 6. "Everything hackable will get hacked." Vercel put $1M on a sandbox-escape challenge. One researcher with a decade of red-team experience, a custom harness, and a mix of open and closed models found a novel Linux kernel vulnerability. Agents are the ideal security researcher, but human skill plus a good harness still wins. 7. Evals are the new marketing. Don't outsource "which model is best" to benchmarks you don't understand. Keep your own traces, reflect on them, derive your own skills, and judge models on your own work. 8. Hiring never stopped, through COVID and AI. He interviews for systems depth (distributed systems, networks, data flow) and agent orchestration. Output measured in changelog entries grows exponentially while headcount grows linearly. 9. On the agent stack: MCP has basically won and is getting more stateless. MCP events and multi-agent collaboration look promising. He thinks identity, wallet and permissioning should belong to the individual, and a cross-harness solution is needed because the future is many agents, not one. Thank you @gakonst, @rauchg and @paradigm for a great session.

Agents are the new users. @vercel now designs for agents, markets to agents, and asks agents what they think. Their DX team is now "agentic developer experience" and its goal for the year is verifiability: linters plus skills, correctness loops. They shipped a site that tells you whether agents can read your website. Most can't. Same is @BlockRunAI , we always put agent as the first citizen.