Warden: Agent Reliability Supervisor

Incidents0

Avg time-to-detectn/a

$ recovered (hard)$0

$ irreversible loss$0

$ loss prevented (est.)$0

Latest benchmark 90 episodes · 3 scenarios × 30 seeds · scripts/bench.py

Detection100%

False-positive (30 healthy)0.00%

Median MTTD1 tick

$ recovered (reversible)$336-$409

$ recovered (refund fraud)$0 honest

03 / Warden live reasoning (via Dynatrace MCP)

05 / Real Dynatrace tenant evidence

The operator console above runs in sim mode so the loop is deterministic and replayable. The same code path also exports OpenTelemetry traces and metrics to a real Dynatrace tenant via OTLP/HTTP. The captures below are screenshots from that live tenant (xwy37883.apps.dynatrace.com), produced by scripts/otel_smoke.py against the @dynatrace-oss/dynatrace-mcp-server. The DQL queries and span attribute schema are documented inline on each card.

service.name = warden service.namespace = warden.fleet OTLP/HTTP → *.live.dynatrace.com DELTA temporality for metrics