Agent tool-wait chip (how long agents are waiting on their tools)

What you are looking at

• The big figure is the middle, not the average. A few very slow calls drag an average up; the middle figure is what a typical call actually waits.
• The small figure is the slow tail. If it is much larger than the middle, a minority of calls are having a bad time even when most are fine.
• It is about the whole machine, not one agent. When many agents run at once, every call waits longer — the same trivial command that takes under a second on an idle box can take ten or more when a hundred agents are working. That is the machine being busy, not a broken agent.
• A dash means nothing was measured. Either no agent has run a tool recently, or there were too few calls to show a number honestly. It never shows zero for "no data".
• Cloud sessions are not counted. The reading is about the work running on this computer.

Why it is cheap

The app already times every tool call it launches, for its own decisions about how fast to start work. The chip only displays that measurement — it starts no probe, reads no file, spawns nothing, and re-reads one in-memory number every fifteen seconds.

Where it lives

It ships pinned to the header. Like every header widget you can hide it, move it, or bring it back from the header settings — and hiding it stays hidden.

What it does not tell you

• Which agent is slow. The reading is the fleet's, not a session's.
• Whether the slowness is the disk, the CPU or a provider. The Performance Monitor panel is where those separate readings live.