Vercel

On 2 January 2026, Guillermo Rauch argued that the CLI is the fundamental abstraction of coding agents, a claim he grounded in their need for OS-level access: they run programs, install missing ones, scan filesystems and read logs. He extended it to a million concurrent Claude Code instances working one bug, then landed on Vercel Sandbox as the compute layer for that load. The move reframes Vercel from web deployment platform to infrastructure for autonomous agents.

Jude Gao published Vercel's agent evals on 27 January 2026, comparing skills (active retrieval) against AGENTS.md (passive context) for teaching agents the Next.js 16 APIs use cache, connection() and forbidden(), absent from model training data. Without updated documentation, agents produced incorrect code in 47% of cases. In 56% of eval runs the agent never invoked the available skill, matching the 53% baseline; explicit trigger instructions raised success to 79%. A static AGENTS.md injected each turn reached 100%, after the team compressed 40 KB of documentation into an 8 KB pipe-delimited index.

Lee Robinson, also at Vercel, argues that AI lets non-developers build one-off tools fitted to their own needs, and that widening who creates software will raise appreciation for well-built software.

The eval result sits awkwardly next to Vercel's seat as a Core Maintainer of the Agent Plugins 1.0.0 specification, alongside Amazon, Cursor, Microsoft and OpenAI: a vendor-neutral format for distributing Agent Skills and MCP servers in one directory. Anthropic, which produced both packaged components, is not among the maintainers.