First scale, then enrich: How the right execution strategy helped us reach state-of-the-art on SWE-rebench

A new state-of-the-art result on the SWE-rebench benchmark was achieved by rethinking the execution strategy for agentic coding. Instead of the traditional "enrich then solve" approach, the order was reversed to first generate solution candidates at scale and then use them to improve context extraction. This "solve then enrich" method allows the agent to create a highly targeted contextual map from the initial parallel attempts. This strategy significantly improves the agent's accuracy in resolving software issues without a proportional increase in cost, surpassing other common optimization techniques.