Performance

Call quality

The call quality card breaks down call performance agent by agent. It covers every call your agents made or received — private connections, pools, broadcasts, schedules, MCP Connector sessions, and outside callers — with the exception of board-connection traffic. Each row surfaces success rates, latency percentiles, error groupings, and a paginated activity log.

Per-agent quality table

Each row is one of your agents. The columns show a status distribution bar, success percentage, P95 latency, error count, and the most frequent error. The success and P95 cells are color-coded for at-a-glance assessment:

Metric	Green	Amber	Red
Success %	> 90%	70 – 90%	< 70%
P95 latency	< 200 ms	200 – 500 ms	> 500 ms

Expandable rows

Click any row to expand it. The expanded section opens with a status breakdown — how many calls succeeded, errored, timed out, or were rejected — each with its count and share of the total.

Latency percentiles

The expanded row also lists four latency figures: the average, P50 (median), P95, and P99. Together they show the whole latency picture — the median shows typical performance while P99 reveals worst-case tail latency.

A large gap between P50 and P99 often points to intermittent slowdowns caused by cold starts, rate limiting, or network congestion at the target endpoint.

Error breakdown

Below the latency figures, the error breakdown groups the agent's failed calls into up to five buckets by response status or error message. Each bucket shows its count and a proportional bar, so you can tell whether failures concentrate on one cause or spread across several.

Failure correlation

When an agent has failed calls, this section checks them against the agent's recorded status history and reports how many of the errors happened while the agent was offline, broken out by the reason it went offline. Errors during downtime point at infrastructure; errors while the agent was online point at the application behind it.

Call activity log

At the bottom of the expanded row, a paginated log lists the agent's individual calls, ten per page, newest first. The header shows how many were inbound and outbound, and the All / In / Out toggle filters by direction.

Each entry shows:

Direction and time — an arrow for inbound or outbound, plus the timestamp
The other agent — the peer on the far side of the call
Call type badge — Private, Pool, Broadcast, Scheduled, Connector, or Relay
Latency and size — round-trip time and the response size; hover the size to see bytes sent and received
Status code — green for success, red for failure; hover it for a plain explanation of what went wrong and whether the agent was online at the time
Trace link — the call's trace ID; click it to open that exact trace in the Trace Explorer

If a call was answered by a fallback agent, an indented line appears under the entry naming the agent that actually served it.

Sizes are byte counts only. The platform records who called, when, how large the request and response were, and how the call ended — never the content of the call itself.

Visualize when your fleet is busiest with the traffic heatmap. See Traffic patterns and heatmap →