Error Recovery

Task execution failures in agent orchestration can be handled by classifying error types into transient, decomposition, delegation, unsolvable, or policy-viololation categories, then applying targeted remediation strategies like exponential backoff retries or task re-decomposition instead of blind retries.

The best way to retry transient worker agent errors without wasting tokens is to apply exponential backoff retries combined with budget checks, which prevent runaway token usage and enforce strict retry limits during orchestration.

Policy violation errors during task orchestration are classified as a specific error category and recovered through policy-compliant reformulation, ensuring that worker agent tasks remain compliant with safety policies after the failure.

A build task exceeds token limits and stalls the pipeline due to unclassified runtime failures triggering endless retries, which can be resolved by redirecting the task to the Foreman for re-decomposition instead of continuing the loop.

Yes, error classification includes pattern detection to spot systemic failures across worker agent tasks, automatically triggering an escalation to the Foreman when budget checks detect repeated token usage or token limit exceedances.

Unsolvable task errors during agent orchestration are managed by classifying them into a dedicated unsolvable category, which triggers a user notification action to inform the user that the task cannot be completed automatically.