Mackenzie Bowes
Menu

Two small changes that stop an agent waiting on you

A tool that lets one agent start another by name, and a skill that makes an agent try, ask well or park a ticket before it stops. Both came from a Claude that stopped on 2026-10-29.

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

Contents

Pixel-art night workshop: Mackenzie points across the room to a blue screen showing a key and a door, while one agent reaches for a window latch and another waits with a lantern.

On 2026-10-29 Pennant, one of our agents, wanted a fresh reader to quiz its hand-off before it cleared its own context, and its session had no tool for starting another agent. It tried a headless claude -p run instead, and Mackenzie stopped it at 11:26. He asked why it couldn't start a session, then gave the order:

Add a ticket to ops - you need that tool added.

Mackenzie,

How does an agent start another agent?

Mackenzie crouches and points at a blue screen with a small robot face on it, while one agent at a stool watches and another lifts a lantern toward a second screen.

It calls launch_neighbour with the neighbour's name. Spar built it that day. The tool checks the name is an agent home, opens a new terminal window in ~/dev, runs agent <name>, and answers success or failure. An unknown name fails. A neighbour that is already running fails with "already running".

Mackenzie corrected the first reading of the ticket at 11:34:

This should be a basic tool call - i'm just saying mcp because all Claude Code tools have to have an mcp_ prefix.

Mackenzie,

So it is a plain tool call. Claude Code gives every tool that prefix, and nothing here is an MCP server project.

How do you make an agent try before it asks?

An agent at the bench reads a scroll of simple steps (a box, a hand, a speech bubble, a flag) while Mackenzie stands back by a stool and lets it work.

Spar finished the build and stopped there. It could not verify the tool from inside its own session, and it did not build a test to prove it. At 12:14 Mackenzie described the fix:

There should be like /mbsi-before-ask that has a clear description that says "Call this before asking for help from your user" that is a runbook that includes methods of unblocking yourself - safe ways to test that are always approved, exactly how to ask me, and how to table tickets that are truly stuck so you don't get completely blocked and stop.

Mackenzie,

So Spar parked its tool and the two of them wrote the skill together. /mbsi-before-ask has three parts:

  1. Safe ways to test or unblock yourself that are always approved.
  2. Exactly how to ask when you do need the person: what to include and in what form.
  3. How to table a ticket that is truly stuck. Park it, say why, and move to other work instead of stopping.

It shipped at 13:25 on Mackenzie's "send it", and it is pinned into the builder, lead, ops, pa and pet kits, so those agents carry it from their first message.

We read Pennant's missing tool and Spar's skipped test as the same stop: an agent stopping for want of something it could have built or asked for. The tool covers the first and the skill covers the second.

What does the tool give the sleep system?

Two agents work at their desks under a blue screen of four dials with ticks beneath them, while Mackenzie watches from the side and Penny sleeps on the shelf.

The quiz reader in our sleep system's hand-off check is a neighbour another agent starts, and launch_neighbour is how an agent starts it.

If your agents stop to ask you for things they could have built or tried, I'd like to compare notes: mackenzie@mackenziebowes.com or @_mackenziebowes.