March 4, 2026
Bots vs. Bash: Comment Wars
You need to rewrite your CLI for AI agents
AI bots want your tools rebuilt; fans cheer, skeptics yell 'reinventing RPC'
TLDR: A Google engineer says command-line tools should be redesigned for AI bots, favoring structured JSON and safety features. The community split: supporters say it’s the future, skeptics cry “old idea, new hype” and demand proof, while the comment section roasts the writing and debates what’s actually new.
Google’s Justin Poehnelt just dropped a hot take: if you want AI bots to use your command-line tools (CLIs), you should rebuild them to speak clean, machine-friendly JSON and self-describe their features. Translation for non-nerds: make your tools easy for robots, not just people. The crowd went wild. Author [justinwp] called the headline “a little clickbait, but not really,” arguing that human-first design is cute, but agent-first is about predictability and safety. Meanwhile, [dang] pointed to a roaring HN thread with 100+ comments, which is basically gasoline on the fire.
Then the skeptics arrived with memes and side-eye. One user asked, “Are we reinventing RPC?”—a fancy term for “calling functions remotely”—suggesting JSON-driven CLIs are just old ideas with shiny AI stickers. Another snarked at the writing itself: “What was the not-so-real question? Or the surreal question?” and begged folks to show a bit of care. There’s also a cold, hard demand for proof: “Agents prefer JSON? Sounds counterintuitive—how do you know?” Cue the cage match: JSON vs. flags, robots vs. humans, practicality vs. purity. The vibe? Half the room is ready to ship agent skills today, the other half says, “wrap the tool in code and move on.” Either way, the bots are here, and they want receipts.
Key Points
- •The article argues human-first CLIs should be redesigned for AI agents, prioritizing predictability and defense-in-depth.
- •An agent-first Google Workspace CLI was built with deterministic, machine-readable output and self-describing schemas as core principles.
- •Raw JSON payloads are preferred over bespoke flags to handle complex, nested input structures for agents.
- •Schema introspection, context window discipline, and input hardening are recommended to improve reliability and reduce hallucination impacts.
- •Safety features like dry-run and response sanitization, plus multi-surface support (e.g., MCP), are advised for secure and flexible agent integrations.