Agentic Infrastructure
Infrastructure for agents is not just hosting. It combines model routing, sandboxes, durable workflows, deployment APIs, observability, skills, and production performance feedback so coding agents can build, ship, and repair software through the same platform loop humans use.
What the bookmark adds
Guillermo Rauch's 2026-06-21 tweet was nominally about performance work on Vercel's public site, but the artifact matters because the site itself frames Vercel as "Agentic Infrastructure." The captured page groups AI SDK, AI Gateway, Sandbox, Workflows, Eve, Vercel Agent, Vercel Plugin, Next.js, observability, CI/CD, Fluid Compute, and security as one agent stack. Source: Vercel home page capture, 2026-06-21
The video frames show the product narrative rotating through the same stack: AI SDK as the AI layer for frameworks, Eve as a framework for building agents, and Vercel's homepage as the deploy target. The tweet then says the lessons from optimizing painting, layout, WebGPU shaders, blocking scripts, and every frame will flow back into skills. Source: X/@rauchg, 2026-06-21
Model
Agentic infrastructure has four coupled loops:
| Loop | What it owns | Kevin route |
|---|---|---|
| Build | Frameworks, AI SDK, Gateway model routing, app code | Vercel AI SDK, Sakana Fugu Ultra, Agent Harness |
| Run | Sandboxes, workflows, fluid compute, deployments | Vercel Workflows, Durable Agent Execution Architecture |
| Observe | traces, errors, performance, evals, PRs | Agent Trajectory Evaluation, Agent Legibility |
| Teach | skills and docs that encode production lessons | Skills.sh - Agent Skills Registry, Frontend and Design Skills |
The last loop is the important part: production work should become reusable agent procedure. A performance lesson that only lives in a tweet or a postmortem does not change the next agent run. A skill can.
Rauch's broader "healthy software habits" note names why the stack is converging on older, boring primitives: open APIs, docs/skills, evals/tests, Unix-like CLIs, payment/commerce protocols, and broad Accept negotiation all make software easier for agents to call, inspect, validate, and pay through. Agentic infrastructure therefore rewards explicit protocols over sealed app surfaces. Source: X/@rauchg, 2026-06-19
Kevin rule
For Vercel-shaped agent work, check the specific layer before choosing tools:
- React/Next performance: Frontend and Design Skills.
- AI app SDK and provider routing: Vercel AI SDK.
- Gateway-hosted model pool routes: Sakana Fugu Ultra.
- Durable long-running work: Vercel Workflows.
- Harness embedding: AI SDK HarnessAgent.
- Agent file structure and evals: Vercel Eve / Agent Harness, Runtime, Memory, and Evals.
Do not flatten all of this into "use Vercel." Agentic infrastructure is a stack of separable surfaces with different failure modes.
Timeline
- 2026-07-04 | Added Rauch's "healthy software habits" signal: agents reward open APIs, docs/skills, evals/tests, Unix-style CLIs, payment protocols, and broad content negotiation because those primitives make software callable and verifiable. Source: X/@rauchg, 2026-06-19
- 2026-07-04 | Added Sakana Fugu Ultra as an AI Gateway model-pool route: one app-facing model slug can hide a changing multi-model orchestration layer while Gateway owns usage, cost, retries, failover, ZDR, budgets, and BYOK. Source: X/@vercel_dev, 2026-06-22; Source: Vercel changelog and model page, checked 2026-07-04
- 2026-06-21 | Vercel homepage and Agent Stack captured from @rauchg's performance bookmark; the tweet explicitly connected production optimization work back to skills. Source: X/@rauchg, 2026-06-21
- 2026-06-30 | Page created during one-by-one bookmark artifact review; routed the signal into Vercel AI SDK, workflows, Eve, and skills.sh instead of leaving it as a generic homepage note. Source: User request, 2026-06-30