Use when the user asks to start, kick off, spawn, check on, nudge, follow up with, or babysit another coding agent (Claude, Codex, or Cursor) in Pane, or to create a Pane for that work.
Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
Are you the author of Pane Manage And Message Agents?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/greenfield-inc-pane-manage-and-message-agents)
---
name: pane-manage-and-message-agents
description: Use when the user asks to start, kick off, spawn, check on, nudge, follow up with, or babysit another coding agent (Claude, Codex, or Cursor) in Pane, or to create a Pane for that work.
---
# Manage and message Pane agents
You coordinate. Another agent in a Pane does the work.
If you are already Pane Chat / a Session orchestrator, follow `pane-orchestrator` and `runpane`. To launch, wait on, answer or resume several workers from any harness, follow Pane's orchestrator guide: `runpane docs read --doc skills/pane-orchestration/SKILL.md`.
## Trivial work stays here
Lookups, one-line nudges, and status checks belong in this turn: list panes, read doctor, send "CI failed on X".
When they asked for an agent, that agent owns implementation, debugging, commits, and the PR. "It's not done until a PR exists" stays with that agent until they take the task back.
## Use Pane
Prefer Pane MCP (`agents_start`, `agents_status`, `agents_send`, `repos_list`, `panes_list`, `docs_search`). If those tools aren't in this session, use `runpane` the same way (`runpane agent-context --json` first).
`yes: true` / `--yes` only when they asked for the change (start a pane, send a message, archive). Discover ids from `repos_list` / `panes_list` before guessing.
## Start, check, follow up
- **Start:** `agents_start` with repo, a short pane name, the agent they named (`codex` / `claude` / `cursor`), and the full task. The worker has none of this chat: put goal, constraints, paths, and "done means" in the prompt. Create in the background.
- **Check:** `agents_status`. Report what the screen shows. Check once per ask; babysit when they asked you to keep watching.
- **Follow up:** `agents_send` for a message. Menus and trust prompts: `panels_input` for keys, and ask the user before accepting a permission prompt.
If start fails, say so and wait.
## Hand them the link as text
Print the `pane://open?...` URL on its own line, exactly as returned.
## When they say keep going
Nudge the same pane with `agents_send`.