In a July 6 interview, Vercel CEO Guillermo Rauch stressed the importance of separating foundation models from agent logic in AI agent development, and made clear the company's stance of providing model-agnostic infrastructure.
AI Infrastructure · Vercel
Splitting Models From Agents: Vercel's Bet on Open Building Blocks
CEO Guillermo Rauch wants foundation models treated as swappable, routable components — the opposite of labs like OpenAI that fuse models and agents. The wager: own the open infrastructure layer where AI-generated code lands.
6M
Deployments per day — half from coding agents
1T+
Tokens/day through the AI Gateway
<20ms
Gateway latency across hundreds of models
Agentic workload is taking over the Gateway
Share of tokens driven by agents — doubled in 6 months.
Eve — "Next.js for agents"
An open-source (Apache 2.0) framework that defines an agent as a directory of files.
agent/instructions.md
•
skills/
•
tools/
durable execution
sandboxed compute
human-in-the-loop
subagents
evals
100+ agents running inside Vercel itself
30K+
questions/month handled by a data-analytics agent on Slack
32×
ROI from a sales agent, at $5,000/year
92%
of support tickets auto-resolved
What developers like
Next.js-style file-based design makes agents easy to build
Production features — durable execution, sandboxing, evals — built in
No token markup, BYOK, failover and observability in the Gateway
Open questions
Heavy reliance on Vercel-native tooling
Hard to self-host or deploy to other clouds
New framework — limited real-world accounts of evals & edge cases
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…