Claude Code 2.139 introduced Agent View, a new interface that lets you see every running and completed session across multiple projects in one place. Instead of spreading work across a pile of terminal tabs and windows, you can manage a large amount of concurrent work from a single terminal tab.
To me, this is a major upgrade in how people interact with coding agents, and its importance is still being underestimated.
This is not a new model and it does not expand the agent’s raw capabilities. What it changes is the way the experience is organized. That turns out to matter a lot. The redesign makes parallel development much more practical. Before this, I had more or less fallen back to a linear workflow, because switching rapidly between multiple tasks was too awkward and too easy to lose track of. After using Agent View, high-concurrency work started to feel natural again.
What it actually fixes
Parallel sessions were possible, but not easy
Running several sessions at once used to mean opening more windows, launching Claude Code again, and starting tasks one by one. That sounds simple, but in practice it is a workflow most people will not comfortably stick with.
Opening a new window feels like a heavy action. And once you send a prompt, the model starts responding almost immediately, which naturally pulls your attention toward that session. It becomes surprisingly hard to ignore the output and move on to opening another window for the next task.
Agent View changes this by making tasks feel asynchronous by default. You enter a prompt, it goes off and runs in the background, and the input box is already there waiting for the next task. There is no active stream of output competing for attention. The transition from one task to the next becomes much smoother. Simply removing the friction of window creation lowers the mental cost significantly and makes it easier to start work in parallel.
Multiple sessions were hard to keep track of
To handle parallel work before, I had even put together a small hook so Claude would ring a bell and show a popup whenever a task finished. Even with that, once several windows were involved, it was still easy to miss updates or forget which window was doing what.
Agent View makes this much easier by grouping sessions by status. You can quickly see which sessions are waiting for your input, which are still running, and which are done. That means you can scan the state of everything from a single place and decide what to do next without constant window switching.
Git worktree isolation used to add too much overhead
It was already possible to isolate concurrent code changes with git worktree, but the workflow was never quite the same as normal development. It also required extra parameters when launching Claude Code, which added yet another layer of mental overhead.
There was another problem: sessions were tied to a worktree, so it was easy to lose track of them.
In Agent View, if a session needs to make changes, worktree isolation is enabled automatically. From the user’s perspective, the worktree is basically invisible. That dramatically lowers the barrier to using it.
Managing several projects at once was awkward
Now you can simply type @ followed by a project name in the input box, and Claude Code will run in that project directory using that directory’s existing Claude Code configuration. There is no need to open another window just to switch projects. That cuts down context-switching cost for the human side of the workflow as well.
The design idea behind it matters
Claude Code also added a set of small interaction details that make Agent View more usable.
Sessions that were not originally started under Agent View can be added into it with the /bg command. You can jump back to Agent View quickly with the left arrow key. Pressing Space shows recent conversation from a session and lets you reply quickly.
The pattern here is worth paying attention to. The goal is clearly to keep the user inside Agent View as much as possible, rather than pushing them into individual sessions.
That makes this more than a management panel layered on top of the old interaction model. It is an attempt to redefine the relationship between user and agent: not as a single chat window, but as a stateful task board.
Current problems
Even so, there are still some rough edges.
- It relies on a daemon process that stays resident, and I have already run into cases where it maxed out CPU usage.
- Each session starts its own Claude Code process, but it is not clear when those processes are supposed to shut down. It is easy to end up with a large number of Claude Code processes consuming system resources.
- The input box behaves differently from the interface inside a session. For example, you cannot
@files there, and you cannot search history. - Slash commands only support skills right now. Other built-in commands, such as checking usage, still require entering the session itself.
- Worktree activation can fail. This may be related to my local git setup; if branch strategy rules are defined in an existing
CLAUDE.md, worktree may not be triggered.
None of these issues change my view of the direction. An interaction model like Agent View is going to become standard. Over time, attention will shift away from managing one session at a time and toward orchestrating more work at once.