toolchurn

noun  /ˈtuːltʃɜːrn/

The cycle of adding and removing agent tools faster than you can evaluate them — where each addition degrades tool selection instead of improving it.

Symptoms

• You connect one more MCP server, and the agent gets worse at a task it already handled.
• Nobody on the team can say which tools are in the context right now.
• Tools get removed because the output "felt noisy" — not because a measurement said so.
• The tool list only ever grows during a demo, and only ever shrinks during an incident.

Why it happens

Tool descriptions compete for the same attention budget. The 40th tool does not cost one-fortieth of the decision — it makes every one of the previous 39 slightly harder to pick correctly. Capability is additive; selection accuracy is not.

Churn compounds the damage: because nothing was measured on the way in, nothing can be measured on the way out. The team is left tuning by vibes on a surface that changes weekly.

What to do about it

• Hold a fixed task set. Run it before and after every tool you add. No number, no merge.
• Set a tool budget. A hard cap forces the real question: what comes out to let this in?
• Retire on evidence. Removing a tool needs the same bar as adding one.
• Log the roster. If you can't reconstruct which tools were live last Tuesday, you can't debug last Tuesday.