Vercel

Vercel CEO Guillermo Rauch wrote on 2 January 2026 that the fundamental abstraction for coding agents is the CLI, a technical necessity rather than a form-factor preference: agents run programs, install missing ones, scan filesystems and read logs. He recast Claude Code as AI for your operating system, projected a million concurrent instances thrown at a single bug, and named Vercel Sandbox as the compute layer for that load.

Later that month Jude Gao published evals run at Vercel on teaching agents the Next.js 16 APIs missing from training data (use cache, connection(), forbidden()). Without documentation, agents wrote incorrect code in 47% of cases. The available skill was never invoked in 56% of evaluation cases, holding success at the 53% baseline; explicit instructions pushed invocation above 95% and success to 79%, though results swung with the exact wording. A static AGENTS.md injected into the system prompt each turn reached 100% on build, lint and tests, and the team compressed 40 KB of documentation into an 8 KB pipe-delimited index while holding that score.

Lee Robinson, also at Vercel, wrote earlier about personal software: AI letting non-developers build single-use tools for their own needs, a shift he expects to raise regard for well-made software.

By 6 August 2026 Vercel sits among the Core Maintainers who published Agent Plugins 1.0.0 with Amazon, Cursor, Microsoft and OpenAI, a vendor-neutral directory format shipping Agent Skills and MCP servers together. Anthropic, which authored both packaged components, is absent from that list.