Unlocking parallel test-time scaling for long-horizon agents (blog.doubleword.ai)

Recent research has unveiled advancements in multi-agent coding environments, specifically enhancing the capabilities of long-horizon agents. A novel inference stack called Doubleword was tested with 64 coding agents, demonstrating the potential for significant computational efficiency by enabling better memory reuse during model calls. In a 24-hour test on 731 programming problems, agents utilizing the Doubleword stack completed all tasks in just over 20 hours, achieving an impressive 51.2% success rate individually and 70.7% collectively when pooling solutions. This contrasts starkly with the throughput-oriented SGLang configuration, which yielded an average of only 24.8 solved problems per agent. The significance of this development lies in its ability to optimize memory handling and computation at scale for AI/ML applications. The Doubleword deployment achieved around 30 times the request and token rates compared to traditional methods while efficiently managing history retention and multiprocessing. By leveraging a more effective affinity-based routing strategy, the system minimized the risk of cache eviction, thus allowing agents to build upon their prior outputs without significant degradation in performance. This breakthrough could lead to accelerated timelines for coding tasks and better resource allocation in coding agents, positioning them as more powerful tools in software development and machine learning workflows.