I was a backend lead at Manus before the Meta acquisition. I’ve spent the last 2 years building AI agents — first at Manus, then on my own open-source agent runtime (Pinix) and agent (agent-clip). Along the way I came to a conclusion that surprised me:

A single run(command="…") tool with Unix-style commands outperforms a catalog of typed function calls.

Here’s what I learned.

Go to Source