GuardedAct: simulating LLM remediation actions in a digital-twin sandbox cut collateral damage from 25.6% to 5.2% for about 8 seconds more recovery time

arXiv 2609.11264·medium signalXBlueskyLinkedInCopy link

The LLM proposes ranked remediation actions for a microservice failure. Each action runs first in a lightweight digital twin that estimates blast radius from live topology and telemetry. A rollback-confidence gate then auto-executes only low-risk actions and sends the rest to a human. On five DeathStarBench fault scenarios it reached 87.4% recovery. This is a concrete shape for any agent with ops write access: simulate the action, classify its blast radius, and auto-run only what can be rolled back.