Scaling Proactive Compute: When Agents Anticipate the Task

Agents often know a lot about a task before it arrives, such as the relevant code repository, filesystems, and documentation. This creates an opportunity for agents to use this context to anticipate the task and prepare for it, spending what we call proactive compute. In this paper, we precisely scale proactive compute by controlling it at the token level using budget forcing, and measure how performance, cost, and test-time efficiency change with it. Across nine benchmarks in software engineering, legal work, algorithmic reasoning, mathematics, and four model families (DeepSeek-V4-Flash, Qwen3.8-27B, gpt-5.6-luna and Muse Spark 1.2), we find that studying for up to 128k tokens substitutes for test-time compute and shifts the accuracy–test-time-compute frontier. On Harvey LAB, a 134M-token legal filesystem, we scale proactive compute beyond a single context window using parallel study agents and compaction. On eight machine-learning engineering tasks, agents that keep working while their commands run reach the median human score 1.6–1.9× sooner. We also find that proactive studying reduces test-time work: on SWE-QA-Pro, a 16k-token study more than halves test-time reasoning and cuts tool calls tenfold at matched performance. In practice, agents that work repeatedly in the same repository or filesystem can benefit from proactive compute by answering faster, giving better answers and using fewer tool calls. More broadly, these results establish proactive compute as another axis of scaling, complementary to test-time compute. Code is available at https://github.com/ScalingIntelligence/ProactiveCompute.