Background sessions and agent view

When there are four things running, the problem stops being how to start them and becomes how to see them.

claude agents

Agent view is one row per session. It shows what is running, what has finished, and, most usefully, which sessions are stuck waiting for a permission answer. You attach to a row to drop into that conversation, and leave it again.

What it is for

Dispatch a bug fix, a review, and an investigation into a flaky test. Go and do something else. Come back to three rows, two finished and one asking a question. That is the shape of work this screen is built for, and it only pays off once the tasks are genuinely independent.

• Independent jobs, not steps of one job.
• Work you are willing to check afterwards rather than watch.
• Tasks small enough to describe in a sentence, because you are not there to steer.

Which raises the obvious question: how do you trust work you did not watch? Everything in part 4 is the answer. Permission rules, a sandbox, and hooks are what make a background session something other than a gamble.

Sessions that talk

Two other features sit near this one and are worth knowing by name. Cross-session messaging lets one session pass findings to another. Agent teams put a lead session in charge of teammates that each work in their own window, which is powerful, experimental, and off unless you set an environment variable to turn it on.