Context-Contaminated Restart Model for LLM Agent Failures

The Context-Contaminated Restart Model (CCRM) is a new formal framework that tackles the issue of LLM agents struggling with multi-step tasks that utilize tools. When these agents attempt a task and fail, they carry the failure into their context window, which negatively impacts subsequent attempts and increases error rates beyond the baseline. The model outlines a sequence of T tool-call steps, each failing at a base rate of epsilon_0; following a failure, the next step operates under a contaminated context with a higher error rate of epsilon_1 > epsilon_0. The paper presents five key findings: a precise formula for the probability of success within K attempts, a cascade-overhead theorem that measures the extra attempts Delta K due to contamination, and an optimal budget-allocation theorem that determines the pipeline depth T* for maximizing success probability. This research, available on arXiv under ID 2605.08563, is the first formal analysis of this commonly observed yet previously unformalized issue.