Agent session (session)

What it is

A session is a single harness process running on a device, tied to one task. It's the concrete, watchable thing: you can drop into a session's live terminal, see what the agent is actually doing right now, and send it input if it needs a nudge, the same way you'd look over a developer's shoulder. When the session ends, whether it finished cleanly or ran into trouble, its full transcript stays attached to the task.

Why it matters to you

When something feels off, a stuck task, a surprising result, a session is where you go to find out why. Being able to watch a session live, rather than waiting for a final result with no visibility into how it got there, is what turns "the agent did something" into an actual explanation. For anyone who's ever had to take a developer's word for what happened, this is the difference between trusting the summary and being able to check it.

How it works in CodeHerder

CodeHerder's Sessions view lets you follow what your agents are doing in real time, drop into a live terminal, and review a finished session's full timeline afterward. The following a live session doc covers watching a session live and sending it auditable input. It also covers managing capacity across queued sessions, including what to do when a session looks stuck rather than simply slow.

• Following a live session →

What people get wrong

A session and a task aren't the same thing. A task can pass through several sessions over its lifetime, one for the code stage, another for review, each a separate run, even though the task itself is the one continuous unit of work moving through the workflow. A session ending isn't the same as a task finishing, either: a session can end because its stage is done and hand off cleanly to the next one, with the task itself still very much in progress overall.