Spaces: A CLI Built for Humans and Agents

Mistral AI has shared insights on building Spaces, an internal command-line interface initially created to help human developers scaffold projects, spin up environments, and deploy code. As the tool grew, the team discovered that coding agents needed a different approach than human developers. Interactive features like TUI module pickers used raw ANSI escape codes that locked agents out, leading the team to decouple information needs from input methods. By ensuring every interactive question has a flag equivalent, such as a -y flag for programmatic execution without stdin blocking, agents gained the ability to operate autonomously end-to-end. Furthermore, Mistral implemented a plugin system where components act as data models, allowing agents to query registries for JSON while humans use traditional pickers. To reduce agent errors and stale assumptions, Spaces automatically generates a structured context.json file and an AGENTS.md rules file on initialization. The developers also eliminated hidden assumptions regarding current working directories and environment variables by relying on explicit parameters with sensible fallbacks. According to Mistral, designing developer tools to accommodate non-human processes naturally improves composability, scriptability, and testability, resulting in a superior tool for both humans and agents. The Spaces CLI was built by Lorenzo Signoretti, Riwa Hoteit, and Sam Fenwick at Mistral AI, with contributions from the Applied AI team.