Toward Risk-Free Deployment of AI Agents

The authors advocate grounding testing and debugging in trajectories (records of reasoning, tool calls, and environment observations) to move toward risk-free deployment of LLM agents. The paper surveys challenges such as the oracle problem, non-determinism, and the lack of sufficiency metrics, as well as debugging paths from automated failure attribution to repair and self-evolution. It presents a deployment-readiness checklist covering the full lifecycle, and notes that formal sufficiency metrics, root-cause attribution over long-horizon trajectories, and the reliability of self-evolving agents remain open problems.