I work at Speakeasy. We build a CLI for SDK generation, and we noticed that features we designed for humans - interactive prompts, spinners, guided workflows - were actively breaking AI coding assistants.
We wrote up what we learned from fixing this:
- Non-interactive flags that bypass prompts entirely
- Structured output modes (JSON, --quiet) to reduce token consumption
- Machine-readable exit codes so agents can determine success without parsing text
The insight: you don't need separate agent-only tools. You need escape hatches that let both humans and agents use the same CLI effectively.
I work at Speakeasy. We build a CLI for SDK generation, and we noticed that features we designed for humans - interactive prompts, spinners, guided workflows - were actively breaking AI coding assistants.
We wrote up what we learned from fixing this:
- Non-interactive flags that bypass prompts entirely
- Structured output modes (JSON, --quiet) to reduce token consumption
- Machine-readable exit codes so agents can determine success without parsing text
The insight: you don't need separate agent-only tools. You need escape hatches that let both humans and agents use the same CLI effectively.
Full post: https://www.speakeasy.com/blog/engineering-agent-friendly-cl...