CLI

Guillermo Rauch, CEO of Vercel, put the thesis in writing on 2 January 2026: the fundamental abstraction of coding agents is the CLI, and his reason is technical rather than aesthetic. He reclassifies coding agents as computer-use agents that execute programs, install missing ones, scan the filesystem and read logs, which requires deep access to the OS layer. His analogy is the terminal editor: nvim runs the same on a Mac and on any machine reachable over ssh, so an agent built on that interface crosses the local/cloud boundary without changing shape. Rauch extends this to a million concurrent instances of Claude Code working a bug, a security sweep, or an issue backlog, then lands the argument on Vercel Sandbox, his own compute product.

Cobus Greyling, Chief Evangelist at Kore.ai, applied the same logic one layer up on 9 March 2026. He reads a repeating collapse: the CLI displacing MCP as the natural agent interface, natural-language prompts in a CLAUDE.md file displacing LangGraph and CrewAI, and now the IDE itself. Greyling takes the five functions an IDE provides, file navigation, editing, syntax awareness, build/run/debug, version control, and shows each one reachable by an agent through text. His inversion is the sharp part: the IDE exists to compensate for human cognitive limits an agent does not have.

Greyling keeps four exceptions, including UI design and compliance workflows. What neither note settles is what a reviewer sees once the visual layer is gone.