When you've got multiple agents running, which one needs you, which is done, which is burning money? The Agents sidebar is the cockpit for triaging them at a glance. This tutorial walks the panel top to bottom — the row anatomy, the status grouping, the peek pane, the dispatch input — and explains the one trust decision you make before you fire anything off.
The chat tab is where you drive an agent turn by turn. The Agent View is where you watch the ones you've stopped driving. Most days you'll have both: a chat in front of you for the thing you're actively thinking about, and three or four dispatched jobs running off to the side — a test suite, a refactor, a snapshot pass — that you'll check on later. The Agents panel is the dashboard for that second pile.
bypassPermissions means before you doAn "agent" in the Agent View is a Claude Code session — but those sessions show up here for three different reasons:
The grouping isn't decorative. Each group answers a different question: "what am I working on?", "what needs me now?", and "what did I do recently?" If you only remember one thing from this tutorial, remember which group answers which question — the rest of the panel hangs off that distinction.
Click the Agents button in the activity bar — the stack-of-rectangles icon, fourth from the top, below Source Control. Hover any activity-bar icon to confirm the label; the Agents one is labeled "Agents."
There's no keyboard shortcut for opening the panel today. If you live in the activity bar, you'll learn the position by muscle memory in a few days.
If you've never dispatched a background agent and don't have a Claude chat tab open, you'll see an empty state: "No active sessions — Open a Claude tab or dispatch a background agent above." That's expected — the panel is dynamic, it doesn't pre-populate anything.
The panel has four stacked regions. Knowing what each one is for makes the rest of the tutorial much easier.
| Region | What's there | What it's for |
|---|---|---|
| Header | "AGENTS" title + refresh button (↻) | Force a re-scan of ~/.lingcode/jobs/ if a job's status looks stale. |
| Dispatch bar | Text field "Dispatch a background agent…" + send button | Type a goal, hit ↩, walk away. See Step 5. |
| Grouped list | Active → status groups → Recent | The roster. Click a row to act on it. |
| Peek pane (when a background row is selected) | Transcript + reply input at the bottom half | Observability without leaving the sidebar. See Step 6. |
Background jobs are grouped by status: Working, Idle, Completed, Failed, Stopped. The grouping makes "what needs me now" obvious — anything under Failed or Working with a long age is what to look at first.
Each background-job row crams five pieces of information into one line. Read left to right, but treat each signal as independent — you rarely need all five at once.
? = needs input. Empty gray circle = Idle. Green ✓ = Completed. Red ! = Failed. Gray stop = Stopped. This is your "do I need to act" channel.$0.0423). Only shown once the job has spent something. Early warning for a runaway loop — if a row's cost climbs faster than you expect, intervene.30s, 5m, 2h, 3d. Tells you how stale the row is — a Working row that hasn't updated in an hour is probably stuck.You don't need to memorize the color codes — hover anything and the tooltip tells you. The point of this section is to know that each glyph means something specific, so you don't tune them out.
Type a complete goal into the dispatch bar — "run the test suite and fix any failing snapshots" — and hit ↩ (or click the send arrow). The agent starts immediately in a detached process, the dispatch field clears, and a new row appears under the Working group within a second or two. You can close the Agents panel, switch projects, even close the window — the job keeps running.
The job runs in the current project's root folder (whatever's open in the editor). If no folder is open, it falls back to your home directory.
Click any row under a status group. The bottom half of the panel splits off into a peek pane, showing:
⏵), tool results (↳ or ✗), and the final result (done or failed)The reply input is where the surprise lives. It's disabled while the agent's process is still alive, and the placeholder text changes based on why:
This is the model: you dispatched an autonomous run. You reply to its conclusions or its pending questions, not to its in-flight thoughts. If you need to interject mid-run, use Stop first (Step 7) and then send a fresh prompt.
Right-click any background-job row (under a status group) for its action menu. Each item has a specific use:
| Action | When to use it |
|---|---|
| Open in Window | The peek pane caps at 12 transcript entries. For the full history, side-by-side with another job, or for a long-running review, pop it out into its own window. |
| Stop | Kill an alive process. Only enabled while the row's process status is alive — after that, the entry reads "Stop (already exited)" and does nothing. |
| Rerun (fresh session) | Re-dispatch the same prompt with a clean conversation. Useful when you want a second-opinion run, or when the first run got confused and you want a do-over instead of a follow-up turn. |
| Pin / Unpin | Pinned rows sort to the top of their group regardless of recency. Use for the two or three jobs you actually care about — keeps the panel scannable when you've dispatched a dozen. |
| Rename… | Auto-generated names from the prompt's first words are often useless ("Run the…", "Fix the…"). Rename to something you can pattern-match at a glance. |
| Delete | Stops the job if it's still running, removes the entry from disk, and discards the transcript. There is a confirmation dialog — read it. |
Concretely, this is what a normal day with the Agents panel looks like:
bash npm test, then read src/<file>.snap, then writes. Close the panel; go back to your main work.$0.0312. PR dot is yellow — checks pending.done with the agent's summary: "Regenerated 4 snapshot files; opened PR #847."The point isn't the specific workflow — it's that you only opened the panel three times, never sat watching the agent work, and the row's glyphs told you everything you needed to know at each check-in. That's what the Agent View is optimized for.
So you don't go hunting for things that aren't there: there is no filter or search box, no multi-select, no bulk Stop/Delete, and no keyboard shortcut for opening the panel. If you have so many jobs that you're missing one of those, dispatch fewer or use Pin to keep the important ones at the top. The "Recent" group caps at 20 — older saved sessions still exist and can be opened from the in-tab history popup inside any Claude Code chat tab.