Introduction
Of the five Monitoring views this is the one a governance reader should care about most, and the one most often skipped. Model calls are what an agent thought. Tool calls are what it did — and doing is where the irreversible consequences are.
Where to find it
AI Governance > Monitoring > Tools.
The tab shows the top 5 tools by run count, with each one's error rate and p50 latency.
Why tool use is the governance signal
An agent that | Worst case |
Says something wrong | Embarrassment, a misled user, a correction |
Calls a tool wrongly | A payment sent, a record deleted, data leaving your boundary |
Both matter. Only one of them cannot be taken back. That asymmetry is why the shipped ISO 42001 pack puts four of its eight controls on the tool call — two on Tool input and two on Tool name — see the ISO 42001 pack.
What this view answers
Which tools your agents use most — The top 5 by run count. The most-called tools are where most of an agent's actions happen.
How often, and whether that changed — A step change in calls to one tool is a behavioural change worth explaining.
Which tools fail — A consistently failing tool is a reliability problem, and often a cost one — failures usually trigger retries.
Whether a tool that should not be called is among the busiest — The most direct governance question on this screen.
Note: An empty Tools view usually means missing instrumentation, not a well-behaved agent. Tool spans require the framework's own tracing; instrumenting only the model client produces complete-looking traces with no tool calls in them. Confirm by opening a trace and looking for a tool step before concluding your agents do not use tools.
The finding worth looking for
Tools shows the top 5 tools only (run count, error rate, p50 latency), so it cannot show unused tools. What it does show is actionable in a way most monitoring findings are not:
A tool called far more than expected is usually a loop, or an agent that has found a way to solve a problem you did not anticipate. Read a trace before deciding which.
A tool called only by one agent is a candidate for narrowing scope rather than granting it fleet-wide.
"The safest tool is the one you did not give it. Every other control is an attempt to recover from having done so."
Turning an observation into a control
Identify a tool that should never be called in some context — Financial transactions, deletions, anything with an external side effect.
Write a blocked-tools control on Tool name — Deterministic, and costs nothing per request — unlike an LLM-based check. See the Blocked Tools check.
Test it before enabling — See Test a Control Before Enabling.
Understand what the tier permits — On an advisory connection the control records rather than prevents — see Where Controls Run.