Observability / Trace record
Every model call leaves a readable trace
One requestRead the exact decision path.
See model choice, cost, latency, risk, and fallback.
Team viewAggregate the same trace fields.
Review spend, incidents, and policy behavior by workload.
Agent runsCompare clients and diagnose harness health.
Agent insights, run waterfalls, and engineering findings in Console.
TuningChange policy with evidence.
Use trace patterns to adjust cost and quality rules.
Trace record / req_8f3k2m9x1p
IndexedModelSonnet
Total cost$0.0182
Latency1.17s
RiskLow
Route reasonGuardrail stateFallback stateBilling link
TimeEventEvidenceStatus
00:00
Gateway receivedOpenAI-compatible request
policy: auto:balanced14 ms00:01
Guardrails passedprompt, PII, tools
standard_policypass00:02
Route selectedquality inside budget cap
claude-sonnet124 ms00:03
Fallback checkedchain ready, not used
gpt-4o-mini -> deepseek-v30 used00:04
Response indexedbilling and audit attached
trace complete1.17sAgent observability
From one request trace to agent-level answers
Multi-step agents share run_id with gateway audit. Console connects waterfalls, spend comparison, and harness engineering diagnostics — no separate APM stack.
Agent insights · Harness health
Console previewAgentSpendHarness
openclaw412 runs
$48.20Watchopencode186 runs
$21.40Healthyhermes-agent128 runs
$16.80Watchn8n93 runs
$12.05ActionHarness healthEngineering diagnostics
WatchCache hit rate42%
Tool definition share31%
Effective steps58%
Tail idle runs22%
Cost decomposition$48.20 attributed
- Compare
- Rank agents by spend, security events, and per-run cost in one table.
- Diagnose
- Surface cache misses, tool schema tax, context growth, and idle tail steps.
- Drill down
- Open any run waterfall with harness notes on highlighted steps.
Operational jobs
Use the trace to answer the next team question.
The same record works for a developer debugging a call and an admin reviewing model spend.
When an answer looks wrongCheck whether model choice, guardrails, or fallback changed the result.
When spend risesAttribute cost to policies, request ids, and workloads.
When a provider failsSee whether fallback was ready, used, or skipped.
When policy feels offCompare route reasons across similar requests before tuning.
When agent spend divergesCompare client profiles by spend, events, and harness verdict in Agent insights.
When runs feel expensiveReview cache hits, tool schema share, and idle tail steps in Harness health.
Make every model decision explainable
Trace route choice, spend, latency, risk, and fallback from one gateway record.