SWE-Router Optimizes LLM Cost for Agentic Software Tasks

Multi-turn agentic Large Language Model (LLM) systems are transforming software engineering, but routing every task to the most powerful (and expensive) frontier models can be inefficient. Existing LLM routers typically make decisions based solely on the initial task description, which often lacks the nuance to distinguish between simple fixes and complex refactors. Researchers have developed SWE-Router, a novel value-based temporal routing approach. Instead of making an upfront decision, SWE-Router allows a less expensive LLM to run for a few exploratory turns. By analyzing the resulting partial trajectory, the system can then decide whether to continue with the cheaper model or escalate the task to a more capable, expensive model if the problem proves more complex. A Bayes-optimality theorem supports that conditioning on partial trajectories never harms routing and is strictly better when exploration provides informative signals. Across various weak and strong LLM pairs, SWE-Router demonstrated significant improvements in cost efficiency for software engineering tasks, while largely preserving the performance levels of the stronger, more expensive models. The researchers also released a multi-LLM trajectory dataset to facilitate reproduction and further research.