Google Workspace engineers redesigned their CLI from scratch to prioritize AI agents over humans. The tool rejects traditional command structures, favoring raw JSON payloads and schema introspection to eliminate hallucination risks. Agents receive structured API responses instead of nested flags, with NDJSON pagination enabling incremental processing of large datasets without context window overload. This approach forces explicit field masking (e.g., `--params '{"fields": "id,name"}'`) to prevent token waste from irrelevant data.

The CLI's agent-first design includes runtime documentation queries and input validation layers. Commands like `gws schema sheets.spreadsheets.create` expose live API specifications, replacing static docs vulnerable to version drift. Input hardening rejects invisible characters, double-encoded URLs, and adversarial paths (e.g., rejecting `../.ssh` traversals). These safeguards address agents' tendency to generate syntactically valid but semantically dangerous requests.

Experts note this shift reflects broader AI infrastructure trends: tools must evolve beyond human-centric interfaces as agents become primary API consumers. Google's implementation ships 100+ skill files encoding agent-specific rules — like mandatory `--dry-run` for mutations — that LLMs can't infer from conventional help pages. The CLI also supports Model Context Protocol (MCP) and Gemini extensions, allowing agents to natively integrate commands without shell escaping.

Key innovation: Treating CLI as living documentation and security boundary. By design, agents can't misuse features without explicit schema knowledge, while humans retain convenience flags. This dual-path architecture avoids maintaining separate tools, a critical balance as AI-driven automation scales across enterprise systems.