What agent view actually ships
Anthropic added an "agent view" feature to Claude Code, its terminal-based coding agent. The pitch is straightforward: if you run multiple Claude Code sessions in parallel — one per branch, one per repo, one per task — the dashboard gives you a single pane to see what each session is doing and switch between them without juggling tmux panes or terminal tabs.
The feature targets a specific user: someone who already runs three or more Claude Code instances at once. That's a real workflow. Power users have been spawning agents per git worktree for months, often through custom shell scripts, terminal multiplexers, or wrapper tools like Conductor. Agent view tries to standardize that experience inside the CLI itself instead of leaving each user to roll their own.
What the dashboard shows is roughly:
- The list of active Claude Code sessions on your machine
- Each session's current state — idle, thinking, awaiting input, running a tool
- A way to jump between sessions without leaving the dashboard
- A consolidated view of what each agent is working on
The whole thing lives in your shell. It's a TUI, not a web app, and that's a deliberate choice consistent with how Claude Code positions itself: terminal-native, no separate desktop surface to keep open.
If you've never run more than one Claude Code session at a time, agent view solves a problem you don't have. The feature exists for the subset of users who already coordinate parallel agents and want less context-switching friction. Solo-session users can safely ignore it.
Why the developer response has been muted
The pushback we've seen isn't that the feature is bad. It's that the hard problems with multi-agent workflows live upstream of "which window am I looking at."
Here's what developers actually struggle with when running parallel Claude Code sessions, and where agent view lands on each:
Coordination, not just monitoring. Knowing three agents are working on three branches doesn't tell you whether they're stepping on each other's changes, duplicating work, or producing diffs that will conflict at merge time. Agent view surfaces state. It doesn't reconcile output.
Cost visibility. Running four agents in parallel multiplies your token spend by four. The dashboard doesn't surface per-session cost in real time, and that's the metric most teams care about before they let agents fan out across a codebase. You have to wait for the Anthropic usage page to update, then attribute spend back to sessions yourself.
Output quality across sessions. When you run multiple agents on related tasks — say, three different approaches to the same refactor — you want comparative review. Agent view shows what each is doing, not how to pick between them. The "diff A vs diff B, here's which I'd merge" view doesn't exist.
Existing tooling already covers the basics. tmux, screen, Zellij, kitty, WezTerm, even plain iTerm tabs handle session management. Developers who already built workflows around those tools see agent view as a reimplementation of capabilities they already have, with less flexibility than their custom setup.
The lukewarm response isn't "this is bad." It's "this isn't the bottleneck." The dashboard solves the easy problem (where are my sessions?) while leaving the hard ones (do these agents agree? what did this run cost? which output should I keep?) untouched.
There's a secondary critique that comes up too: a TUI dashboard in 2026 feels like a half-step. If you're already going to build session orchestration, why not a proper web UI that survives terminal crashes, supports remote sessions, and shows real cost graphs? Anthropic's answer appears to be that Claude Code's identity is CLI-first — but for power users juggling multiple machines or remote work, that constraint is the constraint.
What would actually move the needle
The more interesting question isn't whether agent view is useful — for some users it will be a small quality-of-life win — but what a real multi-agent control plane would look like. Based on what we've seen production teams ask for:
- Per-session cost tracking with thresholds. "Stop this session if it crosses 50k tokens" beats a status indicator. "Show me which session burned 80% of today's spend" beats a list of session names.
-
Conflict detection at the diff level. If two parallel sessions both touch
src/auth/login.ts, surface that before you waste compute on both. Even a static pass over each session's working tree would help. - Output comparison. A side-by-side view of multiple agents' diffs on the same task, with a pick-and-merge workflow. This is the feature that would make parallel agents genuinely productive instead of merely possible.
- Persistent session state. Currently, killing your terminal effectively kills the session. A real control plane would let sessions survive terminal crashes, machine sleep, and laptop close — and ideally let you resume from another device.
None of those are in agent view today. Some are roadmap candidates, some are out of scope for a CLI tool entirely. The honest read: Anthropic shipped maybe 20% of the multi-agent UX problem, and the remaining 80% is still open — across every agent CLI on the market, not just this one.
For developers who want to keep using Claude Code, agent view is a no-cost upgrade. Turn it on, see if it helps your workflow, ignore it if it doesn't. The bigger question — how do you actually manage three or more AI agents on the same codebase without thrashing yourself or your token budget? — remains unsolved across the industry. Cursor, Cline, Aider, Continue, and Claude Code are each working on different slices of it.
Don't let agent view's existence trick you into running more parallel agents than you can review. Token spend scales linearly with sessions; review attention does not. The bottleneck is rarely how many agents you can spawn — it's how many outputs you can sanity-check before merging.
When agent view is worth turning on
Short answer: when you already run three or more Claude Code sessions at once and your bottleneck is genuinely "I can't keep track of which one is doing what." If that describes you, the dashboard will save you small amounts of friction many times per day, and that adds up.
If your bottleneck is anywhere else on the list above — cost, conflicts, review, persistence — agent view won't fix it, and you should keep building around it with the tools you already have. The dashboard isn't the answer to multi-agent coding. It's a session switcher, and judging it as anything more than that sets up the disappointment that's been showing up in the response so far.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)