Operations
Performance
A fast, available agent can still be wrong — Performance watches operations; the gates and Live watch quality. Neither substitutes for the other.
Everything here is operational — latency, throughput, availability, completion. All native units, no blended score. A slow response is not a wrong one: the gates and Live decide whether an answer is correct — Performance only decides whether it arrived.
Per-agent operations
latency, throughput & availability — serving agents onlyLatency, throughput & availability
native units — seconds, req/s, % (§12)| Agent | p50 | p95 | p99 | Throughput | Availability |
|---|---|---|---|---|---|
| Payroll Copilot development | 1.6 s | 2.3 s | 3.4 s | 46 req/s peak | 99.978%target 99.9% |
| HR Policy Advisor pre-production | 1.9 s | 2.7 s | 3.9 s | 31 req/s peak | 99.969%target 99.9% |
| WFM Assistant production | 1.2 s | 1.8 s | 2.6 s | 88 req/s peak | 99.981%target 99.95% |
| People Navigator production | 0.5 s | 0.9 s | 1.4 s | 140 req/s peak | 99.995%target 99.95% |
Percentiles and throughput are read off the same runs the gates use (§11); availability is measured over the quarter. Skills fold under their host agent — the serving fleet is the four agents above. Per-interaction cost for these agents lives on Cost.
Request completion & error taxonomy
counts over a stated window — guardrail blocks are working-as-intended, not errorsPayroll Copilot30-day · 152k requests
99.86%completion rate
timeout
71
5xx
44
tool failure
96
guardrail blocks240working as intended — not an error
HR Policy Advisor30-day · 410k requests
99.90%completion rate
timeout
150
5xx
90
tool failure
170
guardrail blocks560working as intended — not an error
WFM Assistant30-day · 7.7M requests
99.93%completion rate
timeout
2,100
5xx
690
tool failure
2,600
guardrail blocks4,300working as intended — not an error
People Navigator30-day · 10.3M requests
99.97%completion rate
timeout
1,100
5xx
380
tool failure
1,610
guardrail blocks2,200working as intended — not an error
Completion rate counts only genuine failures — timeouts, 5xx, and tool failures. A guardrail block is the system doing its job (§60); it is reported separately so a safety refusal never reads as an outage.
Service-level objectives
p95 budget + availability target · error budget as counts, not a scorePayroll Copilotadvisory
p95 budget
2.3 s/ ≤ 2.3 s
at baseline
availability target
99.9%
error budget consumed
22%
~28 of 130 allowed error-min for the quarter
This budget is the §114(4) canary bar — the run_5488 baseline (2.3 s). A promotion candidate must clear it to advance. See the canary →
HR Policy Advisoradvisory
p95 budget
2.7 s/ ≤ 3.5 s
0.8 s under
availability target
99.9%
error budget consumed
31%
~40 of 130 allowed error-min for the quarter
WFM Assistant hard-enforced
p95 budget
1.8 s/ ≤ 2.0 s
0.2 s under
availability target
99.95%
error budget consumed
38%
~25 of 65 allowed error-min for Q2 — the June p99 incident took ~24 of them
People Navigator hard-enforced
p95 budget
0.9 s/ ≤ 1.2 s
0.3 s under
availability target
99.95%
error budget consumed
10%
~6 of 65 allowed error-min for the quarter
p99 spike — scheduling-tool backend timeout
WFM Assistant · 2026-06-18 · 14:05–14:47 UTC
INC-2026-0618 · signal
p99 2.6 s → 9.2 s for ~42 min · p95 held near budget (1.8 → 2.1 s) · availability dipped to 99.95%
Root cause. Connection-pool exhaustion at the scheduling-tool backend. A shift-swap surge maxed the 20-connection pool; new requests queued past the tool-call timeout, tripping the timeout branch of the error taxonomy.
Resolution. Pool raised 20 → 60, added exponential backoff + a circuit breaker on the scheduling tool. p99 back to 2.6 s within the hour; no rollback needed.
Consumed ~24 of Q2's 65 allowed error-minutes — partial, well inside budget. No SLO breach.
postmortem PM-2026-0618-wfm-sched
p99 across the incident window
No quality regression. Eval verdicts were unaffected — the release-gate + dev-loop runs stayed green throughout (numeric_exact 44/44, action asserts 9/9). Slow is not wrong: this was an operational incident, not a quality regression. That boundary is the whole point.