Writingweeknote · 18 January 2026 · 6 min read

Writing index

When prompts replace pull requests

Vercel's AI agent ecosystem push, generative UI patterns, Claude Code workflows, and MCP Tool Search solving context pollution

On this page

This week painted a clear picture of what AI Engineering is shaping up to look like in 2026. The pace of change is relentless, with Vercel leading the charge on multiple fronts while the agentic development workflow continues to mature.

Vercel is Everywhere

Vercel had an extraordinary week of releases that collectively define the emerging AI engineering stack. They shipped Agent Skills—essentially "npm for AI agents"—packaging a decade of React and Next.js best practices into installable capabilities for coding agents. Alongside this came Agent Browser (faster than Playwright, purpose-built for AI automation), AI Elements Voice components, and updates to the Workflow DevKit with durable execution and improved observability.

But the bigger story is their push into Generative UI. The JSON-to-UI pattern is crystallizing: AI outputs constrained JSON, which renders deterministically through a defined component catalog. This could fundamentally change how we think about frontends—interfaces that adapt to each user, generated on the fly. Rauch's vision of fully generative interfaces feels less like science fiction and more like a near-term roadmap.

The "holy trinity" of agentic UI design resources also emerged this week: rams.ai, ui-skills.com, and Vercel's own design guidelines (now installable as agent skills). Vercel isn't just building tools—they're building the ecosystem.

Claude Code & Ralph Loops Are Changing Delivery

The software delivery process continues to evolve around agentic workflows. Claude Code and Ralph loops are becoming the default way to ship code, with humans increasingly acting as reviewers and orchestrators rather than authors.

This shift is spawning new patterns: markdown plans accompanying pull requests to preserve intent, and the more radical concept of "prompt requests" replacing pull requests entirely. Maintainers receive prompts they can re-execute rather than diffs to review. When AI writes the code, the prompt becomes the artifact that matters.

Rauch's call to return to engineering fundamentals (Unix, CLIs, tests, types, markdown) resonates here. These are the primitives that work well for both humans and AI agents. The tooling is converging on simplicity.

MCP Context Issues: Solved?

A quiet but significant development: MCP Tool Search addresses the context pollution problem that's been limiting MCP adoption. Previously, connecting multiple MCP servers flooded context windows with tool definitions. The new tool search capability enables connecting to hundreds of servers without overwhelming the model—the infrastructure is finally catching up to the ambition.

This feels like a turning point for the MCP ecosystem. The ability to dynamically discover and use tools from a massive library of servers opens up workflows that weren't practical before.

What I was reading

Production AI Agents

Vercel's AI Ecosystem Push

Generative UI

Development Workflow Evolution

Tools and Techniques

Announcements

UI/UX Resources

What I worked on

Food Planning & Cooking App

Finally tackled a long-time pain point: meal planning and cooking workflow. Started with v0.app which proved to be an inspiring jumping-off point. It made a great effort at one-shot of the initial app from a simple prompt and got surprisingly far. The credit model ran dry faster than expected though.

Transitioned to Claude Code and had a working app within a few hours. The v0 → Claude Code handoff is becoming a pattern: use v0 for rapid visual prototyping, then move to Claude Code for the deeper implementation work.

The bulk of the week went into integrating Vercel Voice & Persona for hands-free cooking—the idea being you can ask the app to read out instructions while your hands are covered in flour. Getting AI voice generation working was straightforward enough.

Safari on iOS is where things got painful. The Web Speech API support is inconsistent at best, and the combination of voice input + output + keeping the app responsive while cooking proved frustrating. This continues into next week. The irony of struggling with mobile browser support while reading about Vercel's voice components isn't lost on me.

What I learned

  1. Vercel is building the AI agent ecosystem - Skills as "npm for agents", design guidelines as installable skills, Agent Browser for automation, Voice components. They're betting heavily on agents being the next platform.
  2. Generative UI is gaining momentum - The JSON-to-UI pattern seems to be crystallizing. Define component catalogs, have AI output constrained JSON, render deterministically. Could fundamentally change how we think about frontends.
  3. Code review is being reinvented - Markdown plans alongside PRs, "prompt requests" instead of pull requests. When AI writes the code, the prompt becomes the artifact that matters.
  4. Multi-model verification - Using a "council" of models to review outputs adds robustness. Cross-provider validation as a pattern.
  5. MCP ecosystem maturing - Tool search solving context pollution, ability to connect hundreds of servers. The infrastructure is catching up to the ambition.

What I'm thinking about

  • The Vercel "skills" concept feels significant. If agents become the primary way we interact with dev tools, a package manager for agent capabilities could be as important as npm was for JavaScript.
  • Generative UI raises questions about testing, accessibility, consistency. How do you QA an interface that's different for every user?
  • The return to fundamentals (@rauchg) resonates. Markdown, CLI, types, tests. The things that work well with both humans and AI.

Links

Tools & Frameworks

API & Infrastructure

Design Resources