Quantifying Correction Recovery and Robustness Under Prompt Perturbation in Language Model Agents

Abstract

Language model agents that call external tools now carry out multi-step tasks that change real state, booking rooms, scheduling meetings, placing orders. In practice, people revise their intent partway through, correcting a value they gave earlier. Whether the agent then recovers is rarely measured, even if a failed recovery can leave a wrong action committed while the agent reports success. In this work, we propose a framework to measure what happens after a user correction. Every task is scripted for the agent to know what the correct final state is before it runs, so recovery can be measured against the committed tool state rather than the text the agent is writing. That single choice is what makes the measurement possible. The framework defines four quantities: recovery success, recovery cost, an overcorrection rate, and a silent inconsistency rate that counts runs where the agent says a fix never happened. Four hypotheses are fixed in advance. Recovery cost rises when the correction arrives later. Cost rises when the wrong value is already committed rather than just examined. A significant percentage of the seemingly successful corrections are quietly wrong. Agents more sensitive to small prompt changes are worse off. The framework is built on Voker, an agent analytics platform whose SDK automatically detects user corrections in production agent conversations, so the measurements defined here are meant to run on live traffic as well as on scripted runs. The study is written as a preregistration with method and predictions fixed before data collection, and results are available as a working harness on recovery of correction in tool calling agents.