Agents Are a New Kind of User
The author's insight: CLI tools have been optimized for humans for decades—concise prompts, colored highlights, interactive confirmation, errors humans can guess at. But increasingly the callers are AI agents, and their needs are exactly the opposite—structured output (JSON rather than a human-readable table), explicit machine-readable error codes, an unambiguous non-interactive mode, and help docs detailed to the point of verbosity. Human-friendly design is sometimes an obstacle for agents, and vice versa. Once a tool has two classes of users with wildly different natures, the design has to fork.
The Adaptation Wave in the Tool Ecosystem
This idea is becoming an actual trend: more and more tools are starting to offer an "agent-friendly mode" or a dedicated machine-readable interface, and the rise of standards like the Model Context Protocol (MCP) is essentially about establishing a proper way for tools and agents to converse. For tool authors, this means treating "will be called by agents" as a first-class scenario in design, providing clean structured output, stable interfaces, and clear capability declarations. For the ecosystem, it's a quiet but far-reaching restructuring: software is no longer only for humans, and "for AI to use" is becoming an independent design dimension. Tools that adapt early will be adopted by the agent ecosystem sooner.
via: Hacker News