TACIT-Switch Optimizes LLM Agent Cost with Smart Model Handoffs

Large Language Model (LLM) agents face a trade-off between cost and reliability: smaller models are cheaper but prone to failure, while larger models are more reliable but expensive. The challenge lies in intelligently routing tasks to the appropriate model. Current routing methods often make decisions before execution, after a fixed prefix, or locally at individual steps. This research introduces TACIT-SWITCH, a novel approach that learns permanent handoff policies by leveraging accumulated trajectory evidence and Teacher-Annotated Censored Intervention Times (TACIT). Each annotation is treated as an interval-censored observation on a cumulative-risk scale. The system then employs a mixture-cure threshold model to estimate the probability of a successful strong rollout and, if successful, the optimal handoff threshold, eliminating the need for a teacher during deployment. In multi-step simulations, TACIT-SWITCH significantly boosts success rates by 7.4-11.1 percentage points over existing routing baselines, all while maintaining comparable costs. Ablation studies confirm that task features and cumulative trajectory risk provide complementary information for these decisions. When tested on ALFWorld and DABench, TACIT-SWITCH achieved the highest held-out success among learned policies, demonstrating its effectiveness in balancing cost and performance.