Agent Dashboard & Multi-Session

Use several Grok chats at once—and keep them organized with names and a dashboard overview.

After your first shipAdvanced—skip until you need parallel chats
Placement

This is an advanced lesson. If you haven’t shipped Star Clicker Arena or Focus List yet, do that first—one solid session beats five half-finished ones.

Why multi-session work exists

One chat has a limited context window. Parallel work (research in A, implement in B, review in C) is cleaner as separate sessions than one mega-thread. Multi-session tools help you overview and dispatch without losing the plot.

Agent Dashboard

The Agent Dashboard is a control surface for sessions: see what’s running, jump between chats, queue replies, and start new work (sometimes in a git worktree).

  • How to open it (try in order): type /dashboard (aliases /sessions and /agents-dashboard—same command); press Ctrl+\; open the command palette with Ctrl+P and search for “dashboard”; or from a normal terminal run grok dashboard. Hidden in minimal mode—switch with /fullscreen if the command is missing. Disable with GROK_AGENT_DASHBOARD=0 or [dashboard].enabled = false. F3 / /resume is the past session picker (on disk), not this live board. Headless grok -p sessions appear in that picker without mixing into default interactive history.
  • Not the same as /config-agents (alias /agents), which edits agent definitions and personas—not the live multi-session board.
  • Use it when you have more than one active chat and need an overview. Rows can show a short summary of the previous turn. From the dashboard, Shift+Tab can start a new agent in Auto mode (fewer permission prompts)—stay on Ask while learning.
  • Always /rename sessions so the list is readable (example: “combat-slice”, “ci-fix”). From the dashboard you can also rename with Ctrl+R.
  • Delete a session: while inside a chat, /delete (confirm)—if you opened that chat from the dashboard, you return to the dashboard. On the roster, Ctrl+X twice within 2s, or hover and confirm [✗].
What success looks like

A list or board of top-level sessions with statuses (needs input, working, idle, …). If none of the open steps work: that feature may be missing or disabled on your build. Use two terminal windows instead, and /rename each session so you can tell them apart.

Patterns that work

Slice isolation

One session per vertical slice or journey step. New session when the topic flips.

Explore // implement

Session A: read-only mapping. Session B: edits with tests. Link findings via a short NOTES.md in the repo.

Worktree experiments

grok --worktree=feat "…" keeps risky branches off your main checkout.

Background inside a session

Long builds use background tasks (Workflows lesson); dashboards manage whole sessions.

Hygiene rules

  • Rename sessions at milestones
  • Write decisions into the repo (AGENTS.md / DECISIONS.md)—don’t rely on five open chats as memory
  • Compact or close dead sessions
  • Never Always-approve on untrusted worktrees

Lab

  1. Open two terminals in the same project (or use dashboard if available).
  2. Session A: “List the top-level files only; do not edit.”
  3. Session B: “Add a one-line comment to README if it exists; if not, say so.”
  4. /rename both sessions meaningfully.
Done when

You can run two focused sessions without mixing goals, and you know where to look for a multi-session overview on your build.