Unlocking parallel test-time scaling for long-horizon agents
a day ago
- Repeated history in long-horizon coding agents creates opportunities for computation reuse, but parallel agents compete for memory, affecting throughput.
- In a benchmark of 64 agents on 731 SWE-bench Pro problems, the Doubleword inference stack completed all 46,784 attempts in 20h23m, outperforming SGLang deployments that averaged 10.7 and 24.8 problems per agent in 24 hours.
- Doubleword achieved 30x higher request and token rates over throughput-optimized SGLang, mainly through 98.1% prefix cache reuse (vs 2.16%) due to affinity routing and KV cache retention in host memory.
- Higher throughput enabled 64 attempts per problem, raising resolution success from 51.2% (single agent) to 70.7% (any of 64 agents), a 19.5-point improvement.
- Cost per attempt was $0.049 with Doubleword, much lower than $3.92 (SGLang baseline) and $1.70 (throughput-optimized SGLang), making repeated sampling economically viable.
- DeepSeek-V4-Pro with 64 attempts per problem matched or exceeded proprietary models (Gemini 3.1, GPT-5.6, Claude Opus 4.8) at similar inference spend, demonstrating parallel test-time scaling for open-weight models.