Fleet OverviewAttention

2 critical regressions and 3 SLA risks need investigation

Fleet is 99% healthy, down from 100% previously. 2 critical regressions detected. 3 agents trending toward SLA breach. $3340 projected overspend.

Investigate regressions Review SLA risks Open budgets

Fleet health

Group concentration by

Risk concentration

Where active risk is concentrated before jumping into traces or regressions.

Risk posture trend band

Is reliability, regression pressure, and budget pressure moving together or independently? Read the band first, then escalate.

Reliability posture

Open traces

Regression pressure

1.67

Review regressions

Budget pressure

1670

Open budgets

Decisions queue

Issue · Insight · Action

ACTIONTighten the refund-policy prompt against stale-context answers

Proposed prompt edit closes the cluster Foxhound just surfaced. Approve, edit, or reject — the diff applies in your repo, not in Foxhound.

Claude Agent SDKAdd a "use only provided refund-policy context" guard to the system prompt.

Expected impact: faithful-to-context evaluator score expected to lift from 0.71 → ≥ 0.92 within 100 traces.

Open prompt Compare runs

ISSUEInvestigate 2 critical regressions

The refund incident narrative is still the most important demo anchor. Start with the seeded regression path, then move into trace, diff, replay, and prompt evidence.

Regression Trace Run Diff Replay

ISSUEReview 3 SLA risks

Seeded carrier and orchestration paths still show enough latency pressure to justify SLA review before calling the demo fleet stable.

SLAs Traces

ISSUEContain $3340 projected overspend

Refund-policy-agent overspend remains part of the seeded executive story, so cost review should stay tied to experiments and budgets together.

Budgets Experiments

INSIGHTFailure cluster: stale-context answers in refund-policy flow

47 traces failed the faithful-to-context evaluator in the last 24h. Retrieval succeeded; the LLM answered from prior knowledge. This is a prompt issue, not a retrieval issue.

Open evaluator Trace cluster

INSIGHTValidate promotion-ready candidates

The recovery candidate is ready for review with trace-derived experiment evidence and prompt history available from the same seeded scenario family.

Experiments Prompt Notifications

Executive summary