Microsoft 提出 BOXBOX：让智能体在长时间运行中自我改进 harness
微软这篇论文解决了做长时程智能体的人都会碰到的痛点：跑几小时的任务，harness 不合适也没法中途换。BOXBOX 支持运行中热重载，在 PerfRace 上比最好基线多拿到 1.70x 对 1.54x 的加速。写智能体框架或跑长任务的开发者值得看。
Microsoft 与合作者发布论文，提出 BOXBOX 方法：在数小时的长任务运行中，允许智能体自行修改自己的 harness（执行框架），并由此提升长时程任务表现。
传统 harness 优化是离线进行的：在带参考答案的任务上调优后复用。但开放式的长任务没有参考答案，且会遇到调优时没见过的问题。BOXBOX 把优化循环搬进任务本身——智能体自主决定何时修订 harness，内核会热重载新版本并保留完整执行上下文，运行无需重启即可继续。
研究团队在新基准 PerfRace 上测试，该基准要求智能体花费数小时加速真实软件项目。在相同时间和成本预算下，BOXBOX 在四小时时达到 1.70x 加速，而最佳基线为 1.54x。
Great work on harness self-improvement from Microsoft and colleagues.
It discusses how to build long-running agents that improve their own harness,
It turns out that if you run agents for hours, giving them a way to fix their own harness during the run improves results on long-running problems.
More from the paper:
Harness optimization usually runs offline.
You tune the harness on tasks that have reference answers, then reuse it. Long open-ended tasks have no reference answers, and they run into problems the tuned harness never saw.
BOXBOX moves that loop into the task itself.
The agent decides when to revise its harness, and a kernel hot-reloads the new version while keeping the full execution context, so the run continues without a restart.
They test it on PerfRace, a new benchmark where agents spend hours speeding up real software projects. Under the same time and cost budget, BOXBOX reaches a 1.70x speedup at four hours, against 1.54x for the best baseline.
https://t.co/UQgvVqgSQQ
Chat with Paper: https://t.co/9NGxXIpJ8f
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include BOXBOX, Microsoft, 智能体, PerfRace.
Source attribution: curated from DAIR.AI (@dair_ai).
