Two agents, compared honestly

The overlap is most of it

Enough that the practice is learned once and carried across.

1. They run where your code is: read, edit, execute, read the result, continue.
2. They ask permission for some actions, and where that line sits is configurable.
3. They read a file of project instructions from the repository root.

Axes worth comparing

Rather than a feature table that ages badly, the questions whose answers change how a day of work feels.

1. Default autonomy: how many steps before it comes back to you.
2. Permission model: what it asks about, and how granular your answer can be.
3. How changes surface: a diff before, after, or only if you go looking.
4. Session handling: what happens when the working memory fills up.
5. Where it runs: terminal, editor, a machine that is not yours.
6. Licensing and cost model, which differ and are worth checking at the source.

What this page will not tell you

No commands, flags, model names, context sizes, prices or limits. Those change on a schedule measured in weeks.