Today I have learned: with Claude Managed Agents, do not attempt to define them in the UI.
There's a cli for that, and it offers options that don't exist in the UI. And of course, you can make claude code work it for you
To get aligned with agents, I make them build me a fancy HTML description of what they're thinking. Then I tweak it and comment on it until we both know what we're about.
Here's my iterate-on-docs skill, plus a tool that makes this work with HTML.
