Cloud/AI Governance/Tool Reliability
Tool Reliability
Every tool your AI can call, graded: the slowest ones ranked with their spikes, the failing ones split into Agent, Upstream and Vinkius causes, and a latency versus failure matrix of the whole surface.
A fleet can be healthy while one tool quietly drags every conversation through the mud. Tool Reliability is the screen that ends the guessing: every tool your agents can call, ranked twice, by slowness and by failure, with each failure attributed to a cause. The console's own line: "See which tools are slowing everything down and which ones are failing silently. Ranked, diagnosed, and ready for you to act on."
Tool Reliability
See which tools are slowing everything down and which ones are failing silently. Ranked, diagnosed, and ready for you to act on.
Slowest Tools
Avg latency (amber) · max spike (red)Most Failing Tools
Stacked by Agent / Upstream / VinkiusHealth Matrix
Latency × Failure rate · bubble = call volumeSlowest Tools Detail
| Tool | Calls | API | Total | Max | Rate |
|---|---|---|---|---|---|
| search_documents | 412 | 512ms | 823ms | 1682ms | 2.2% |
| create_issue | 386 | 448ms | 761ms | 1503ms | 1.8% |
| list_channels | 341 | 401ms | 704ms | 1412ms | 1.5% |
| get_user_profile | 298 | 372ms | 655ms | 1341ms | 1% |
| send_notification | 264 | 338ms | 592ms | 1204ms | 0.8% |
| update_task | 231 | 305ms | 531ms | 1086ms | 1.7% |
| query_database | 208 | 271ms | 478ms | 967ms | 0.5% |
| upload_file | 187 | 238ms | 413ms | 854ms | 1.1% |
| delete_record | 156 | 205ms | 352ms | 728ms | 0.6% |
| sync_calendar | 124 | 168ms | 297ms | 601ms | 0% |
Most Failing Tools Detail
| Tool | Calls | Errors | Agent | Upstream | Rate |
|---|---|---|---|---|---|
| search_documents | 412 | 28 | 20 | 7 | 6.8% |
| create_issue | 386 | 24 | 17 | 6 | 6.2% |
| list_channels | 341 | 19 | 13 | 5 | 5.6% |
| get_user_profile | 298 | 15 | 10 | 4 | 4.9% |
| send_notification | 264 | 11 | 8 | 3 | 4.1% |
| update_task | 231 | 9 | 6 | 2 | 3.9% |
| query_database | 208 | 7 | 5 | 2 | 3.4% |
| upload_file | 187 | 5 | 3 | 1 | 2.7% |
| delete_record | 156 | 3 | 2 | 1 | 1.9% |
| sync_calendar | 124 | 0 | 0 | 0 | 0% |
The mockup runs on the console's Sample Data. On a paid plan the same screen grades your live tools, per organization.
The four counters
- Tool Calls with the derived success rate: the health of the whole tool surface in one number.
- Unique Tools active in the period, so you know how wide the graded surface is.
- Avg Latency with the familiar split: the emerald share is the upstream API, the amber share is Vinkius overhead. Tool slowness is almost never the proxy.
- Total Errors, red the moment they exist, with the three way split underneath: Agent (4xx, the caller's fault), Upstream (5xx, the provider) and Vinkius (thrown by the proxy itself).
The two rankings
Slowest Tools ranks tools by average latency, and each bar carries its max spike as a red dot: a tool can average 300ms and still spike past a second at 2 AM, and this chart refuses to hide that. This is the list to bring to whoever owns the Connector.
Most Failing Tools ranks by error count, and every bar is stacked by cause: amber for Agent errors, red for Upstream, violet for Vinkius. The stack tells you whose problem it is before you even open the table. A tool failing mostly on Agent errors usually means a scope or parameter problem on the calling side; mostly Upstream means the provider is flaky; any violet sliver means the proxy intercepted something and said so.
The Health Matrix
The third chart plots every tool as a bubble: latency on one axis, failure rate on the other, bubble size is call volume. The top right corner is the danger zone, a slow tool that fails often; a big bubble there is your next incident, and the matrix is where you see it forming.
The two detail tables
Slowest Tools Detail gives each slow tool its calls, the API / Overhead / Total latency split and the max spike. Most Failing Tools Detail gives each failing tool its error count and rate, broken into the Agent / Upstream / Vinkius columns, so the diagnosis ships with the number. Both tables are clickable into the tool's detail, and both empty states are the console's own: "No tool performance data recorded in this period." and "No tool error data recorded in this period."
From symptom to owner
The Details link on Mission Control's Tool Health Matrix lands here. From a failing tool, the next stop is Request Failures to see the actual requests that broke, and the Connector Policy settings are where a repeatedly abused tool gets throttled. This screen turns "the AI feels slow" into: this tool, this server, this latency, this cause, this owner.