What I want from an agent harness

A running list. Mostly about seeing what is happening and being able to stop it.

Let me stop it mid-flight without losing the work. The most common thing I want is not to cancel, it is to redirect. Right now stopping usually means throwing away the context and starting the conversation again, which makes me let bad runs continue longer than I should.

Show me the token budget as it is being spent. Not a total at the end. I want to know, during a run, which part of the window is going to tool definitions, which to files pulled in, and which to actual reasoning. I keep guessing at this and being wrong.

Make the diff reviewable before it lands. Reading a change after it is written into the working tree is fine for small things and terrible for large ones. I want the same staged review experience I would get from a colleague.

Give me a replay. When a run goes wrong I want to see the exact sequence of inputs and decisions again, not a summary of it. Summaries are written by the same thing that made the mistake.

Let verification be a first-class step, not a convention. If the repository declares a check command, the harness should run it and treat failure as failure, rather than leaving it to me to remember.

Most of these are observability requests, which I did not expect when I started writing the list. What I want is not a smarter agent. It is a cockpit.