Production

Observability

Governance-first observability: a live dashboard of runs, decisions, spend, and guardrail blocks — plus optional OpenTelemetry export. It rides the existing audit pipeline, so there are no runtime changes.

Enable it

observe.py
runtime = Antraft.agent(model, tools, task="...").observe().build()
await runtime.run()

Then run the control plane and open the dashboard:

shell
antraft serve
# http://localhost:8000/observability
Note
The dashboard is governance-first: it shows not just what happened, but what was allowed, denied, paused, or blocked — and the spend behind every run.

What you see

  • Live and completed runs, with status (running / completed / killed).
  • Per-run decision counts: allow / deny / pause.
  • Token usage and dollar cost.
  • Guardrail blocks.
  • Final answers.

Programmatic access

tracer.py
from antraft import get_tracer
 
snap = get_tracer().snapshot()
snap["totals"]   # {"runs": ..., "allow": ..., "deny": ..., "cost_usd": ...}
snap["runs"]     # per-run summaries

OpenTelemetry

If opentelemetry is installed, audit events are also emitted as span events — wire Antraft into your existing tracing backend with no extra code.

Next →
Deployment