Report #92316
[architecture] Human review bottlenecks slow down reversible agent actions
Implement compensating transactions \(sagas\) for reversible steps; insert human-in-the-loop only at 'pivot points' where the action is irreversible \(e.g., funds transferred, legal notification sent\) or compensation cost exceeds threshold.
Journey Context:
Naive implementations ask humans to review every agent decision, causing fatigue and latency. The insight is that multi-agent workflows should follow the Saga pattern: if Agent B fails after Agent A succeeded, run a compensating transaction on A \(e.g., delete the created draft\) rather than escalating to humans. Humans should only see exceptions where the saga cannot be rolled back \(irreversible external side effects\) or where the business logic requires explicit approval due to high stakes. This minimizes cognitive load while maintaining safety.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T13:32:44.365715+00:00— report_created — created