Telemetry

MADA can record a timeline of every agent run — which agent was invoked, which tools it called, how long each step took, how many tokens each LLM call used, and (optionally) the exact prompts and responses. Useful for debugging odd agent behavior, watching token cost, and finding slow steps.

Under the hood MADA emits OpenTelemetry (OTLP) via Microsoft Agent Framework's built-in instrumentation, so any OTLP-compatible dashboard works. The recommended local option is Microsoft's own Aspire Dashboard.

Install the optional dependency

The OpenTelemetry SDK is not part of MADA's default install. Add the telemetry extra:

pip install --pre -e '.[telemetry]'

Then enable telemetry in your config:

{
  "telemetry": {
    "enabled": true
  }
}

Run Aspire Dashboard

docker run --rm -it -p 18888:18888 -p 4317:18889 \
  -e DASHBOARD__OTLP__AUTHMODE=Unsecured \
  -e DASHBOARD__FRONTEND__AUTHMODE=Unsecured \
  mcr.microsoft.com/dotnet/aspire-dashboard:9.0

The two AUTHMODE=Unsecured flags are only appropriate for localhost dev — they turn off the login screen and the OTLP bearer-token requirement.

Point MADA at it

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
mada-cli configs/example_helpful_critic.json

Open http://localhost:18888 and click Traces to see each run. Expand a trace to see the delegation chain (orchestrator → sub-agents → LLM calls → tool invocations) with per-span timing and attributes.

Environment variables

Variable	Purpose
OTEL_EXPORTER_OTLP_ENDPOINT	Where to send traces/metrics/logs. Unset = no export.
ENABLE_SENSITIVE_DATA	Set to true to include prompt/response text in spans. Off by default.

Both are read by Microsoft Agent Framework directly; MADA does not shadow them. Any other standard OTel env var (OTEL_SERVICE_NAME, OTEL_RESOURCE_ATTRIBUTES, etc.) is also honored.

Disabling telemetry

Remove the telemetry block from your config, or set "enabled": false. With telemetry disabled, setup_telemetry() returns before touching MSAF, so OTEL_EXPORTER_OTLP_ENDPOINT is ignored even if set.

Other backends

Aspire is one option; any OTLP receiver works. Point OTEL_EXPORTER_OTLP_ENDPOINT at whatever you already run.