Why not keep one long-lived agent?

Why one task per session

Because context from one task is a liability in the next, and because a fresh session is the only way to be sure it is gone.

The problem with a long-lived agent

An agent that does five tasks in one session carries all five into every decision. That sounds like an advantage and is mostly a cost:

• It gets more expensive every turn. The whole conversation is resent. A session’s cost per turn can end up several times what it started at — you can watch the curve do it.
• It gets more confused, not less. A constraint from task two gets applied to task four. A file read in task one is assumed unchanged in task five, when task three rewrote it.
• A mistake compounds. A wrong assumption early is repeated with growing confidence, because the conversation contains it as established fact.

What Muster does

Each task starts a fresh session, seeded with the task in full — its title, description, how the defect shows up, the acceptance terms, and the paths it may touch. Everything that task needs, and nothing from any other.

When the task ends, so does the session.

What you give up

Warm context. The agent re-reads the files it needs, which costs tokens and a few seconds.

That is the trade, and it is deliberate: paying to re-read a file is cheap and predictable, while a task closed on a stale assumption is expensive and arrives disguised as success.

When you do want continuity

You still have it, by hand: resume a session to keep going with its context, or branch it to explore without polluting the original. What the queue will not do is silently carry context between tasks you told it were separate.