David WalshSubscribe
Research preview · Claude Code 2.1.139 · May 11, 2026

Agent view, live.

Anthropic shipped a real dashboard for parallel Claude Code sessions this week. I spent a day with it, then rebuilt it here — keyboard and all — so you can scrub the workflow, hit the gotchas, and walk away knowing whether it changes how you work.

≈ 90 seconds end-to-end·6 lessons · 13 steps · 1 decision · 0 API calls·by David Walsh, daveisbest.org
Agent view is a desktop terminal tool — this page is best on a wider screen. On mobile, you can still scrub through and read along, but the keyboard layer and floating callouts are hidden.
Chapter Boot·1 / 19
claude · agents · ~/dev/clawd-gamev2.1.139 · supervisor up
$ claude agents5 sessions · 1 repo
supervisor up1 needs youpress ? for shortcuts
Pinned(1)
clawd walk cycleWriting src/animations/walk_cycle.ts (frame 8/16)3m
Needs input(1)
power-up designShould the new power-up be double jump or wall climb?1m
Ready for review(1)
jump physicsPR #314 · all checks green · awaiting review12m
Working(1)
collision detectionEdit src/physics/CollisionSystem.ts (28% of file)4m
/loop(1)
playtest level 3run 12 · all checkpoints cleared · next in 4min 4m 00s
background⇧↵attach@repo / subagent
Chapter · Boot
$ claude agents
00s / 77s3 live1 need you
Space play / pause prev / next chapter12 answer the decisionclick any chapter tick to jump
02 · The lessons, six of them

What I wanted you to feel.

Six things changed how I use Claude Code after a day with agent view. Each chapter maps to a beat in the sim above. Click any card to jump the simulator there.

03 · what I actually think

Three things I changed after a day with it.

I stopped foregrounding everything.The old reflex was to keep Claude in front of me — partly to course-correct, mostly because watching it work felt productive. It wasn't. With agent view, I dispatch first and check in when a row turns amber. The first afternoon I shipped three things I'd been putting off because I “didn't have a clean block of time.” Async beats clean blocks.

I started treating needs_input as my only notification surface. Working rows can work. Failed rows want a postmortem, not a panic. Only amber means you. Once I stopped scanning the whole table every two minutes, I got my attention back.

I'm still skeptical that more visibility makes the agents more reliable. The dashboard solves a coordination problem, not a trust problem. Rob May was right about that. But the coordination problem was real, and it was eating my week. So: useful, not transformative — and a giant quality-of-life win the day it shipped.

04 · keep this open

Seven keys, one paste.

The handful of shortcuts that actually changed my muscle memory, plus a copy-pasteable starter so you don't have to think about the first invocation.

$claude --bg "tighten the hero copy; one paragraph, first person"
SpacePeek the focused row without attaching. Answer multi-choice with number keys.↗ chapter
EnterFrom a row: attach. From the dispatch bar: send to background. ⇧↵ dispatches and attaches.↗ chapter
From any attached session, drop back to agent view. The session keeps working.↗ chapter
@repoTarget a sibling repo from the dispatch bar. Also disambiguates subagents — @reviewer beats bare reviewer.↗ chapter
Ctrl+TPin the focused row to the top group. Reserve pins for long-running work you want to watch.↗ chapter
Ctrl+XStop the focused session. Press again within 2s to delete — also deletes the worktree. Merge or push first.↗ chapter
?Show the full shortcuts panel inside agent view. The real one is longer than this list.↗ chapter