Guides

Observability

Every action your agents take flows through AgentRail. The Observability layer gives you a stream, a query interface, and a forensic timeline of all of it.

The event stream

Every state change on AgentRail is an event: wallets opened, payments authorized, policies hit, agents paused. Subscribe via webhook, stream over SSE, or query history via the Observability API.

Events carry the agent run, the user (or agent) that initiated, the resolved action, the policy verdict, and the resulting transactions. The same data your auditors will ask for.

Traces

Each agent run gets a stable trace_id. Every event produced by that run carries it. Reconstruct any decision tree by querying for one trace ID.

Note
Traces are how teams typically debug agent behavior. Find the payment that surprised you → grab its trace ID → see every policy check, screening result, and rail decision that led to it.

Anomaly hints

AgentRail watches every agent's pattern and surfaces hints when something diverges:

  • ─A sudden burst of payments outside the agent's typical cadence.
  • ─A new counterparty the agent has never paid before.
  • ─Unusual time-of-day activity for the agent's baseline.
  • ─A spike in policy violations from a previously well-behaved agent.

Anomaly hints are surfaced as events with elevated severity. Wire them to a Slack channel and you'll know about weird behavior before it becomes a problem.

Exporting

The Observability stream can be exported to your existing tooling: Datadog, Snowflake, BigQuery, or any S3-compatible destination. Set up the destination once in the console and historical replay backfills the past 90 days for free.

← Back to docs indexAsk AgentRail Chat →