Observability
What your agents did, and what it cost.
Every metric here is aggregated from the append-only audit log, the same record of every thought, tool call, and approval decision.
5
Runs
$1.10
Est. cost
326,470
Tokens
0%
Approval rate
100%
Human override
Actions taken, by risk
Of 48 tool calls, this is how many ran at each risk tier. HIGH calls each passed the approval gate.
Approvals
The human override rate is the signal to watch: a rising denial rate on a tool or agent means the classifier or the agent’s instructions need attention.
Runs per day
08-29
08-30
By agent
AgentRunsEst. costGatedDeniedOverride
Warden Bug Fixer
2
$0.53
1
1
100%
Knowledge Assistant
1
$0.03
0
0
—
Warden Self-Audit
1
$0.42
0
0
—
File Organizer
1
$0.11
0
0
—
By tool
ToolCallsErrorsErr rateAvg latency
read_source LOW
36
0
0%
6 ms
list_source LOW
4
0
0%
9 ms
list_files LOW
3
0
0%
7 ms
run_selfcheck LOW
3
0
0%
128 ms
read_file LOW
2
0
0%
6 ms
Telemetry is governed too. Sensitive fields in tool arguments and results (17 key patterns: tokens, secrets, emails, and more) are redacted before export ✓ on, and the audit log is hash-chained so any edit or deletion is detectable ✓ 4 verified. OpenTelemetry export (each run a trace, each step a span) plugs these into Datadog or Grafana.