Abstract
This paper focuses on the task execution problem of large language model agents under tool uncertainty conditions, proposing a unified strategy framework of risk-aware planning and conservative execution to improve stability, controllability, and consistency in multi-tool collaborative scenarios. The method models the agent interaction process as a belief-based sequential decision-making process, using a belief update mechanism to integrate task state and tool observations, and employing an uncertainty metric to characterize the current cognitive confidence level. Simultaneously, it introduces tool reliability assessment to characterize the availability and consistency of tool invocation online, thus explicitly incorporating risk penalties into the planning objective and prompting the strategy to achieve an adjustable trade-off between benefits and risks. At the execution level, the framework employs a safety constraint and action gating mechanism, jointly using reliability thresholds and uncertainty thresholds as trigger conditions for tool invocation, and activating a backoff action when constraints are not met to limit error propagation and out-of-bounds operations. Comparative experiments show that this method exhibits a more balanced advantage in terms of accuracy, task completion, tool invocation success rate, and constraint violation control, verifying the effectiveness of incorporating risk modeling into planning and co-designing with conservative execution, providing an achievable technical path for the reliable deployment of tool-enhanced agents in complex toolchain environments.
