Reliability and operations¶

An AI task fails in more ways than a SQL task, and it costs money each time it runs.

Durable execution replays the model and tool calls an agent already completed, so a retry pays only for the steps that did not finish.
Retry policies lets a model classify a failure and decide whether a retry is worth it at all, with a plain rule table as the floor.
Observability (OpenTelemetry tracing) exports model and tool calls as OpenTelemetry traces and metrics.
Securing agent tools is the defense-layer guide for agents that hold tools: what an agent can reach, how allowed_tables is enforced, and the production checklist.
Provider fallback fails over to another vendor inside one task attempt when a model provider is down or rate limiting. It is configured on the connection, so it also appears under Supported model providers.
Troubleshooting lists the errors a first Dag most often hits, with the fix for each.