NVIDIA发布SoL-Pi：自动进化Agent框架，Token消耗减半
Build your own harness, folks.
Agent工程落地必看，这套自动进化机制能直接砍掉一半Token开销，实测省钱效果显著，建议关注其开源实现。
Build your own harness, folks.
自己动手打造工具链，各位。
This is absolute banger paper from NVIDIA on self-evolving agent harnesses.
这是 NVIDIA 关于自演化智能体工具链的一篇重磅论文。
(bookmark it)
（建议收藏）
They introduce SoL-Pi which cuts token traffic by nearly half.
他们提出了 SoL-Pi，可将 Token 流量削减近一半。
And it matches its baseline harness on GPT-5.6 Sol and Opus 5.
它在 GPT-5.6 Sol 和 Opus 5 上达到了与基线工具链相当的性能。
More details below:
更多细节如下：
Instead of tuning a harness by hand, they run auto-research loops at the harness layer across many repository-derived and verifier-driven environments, keeping only the mechanisms that survive selection.
与其手动调优工具链，不如在工具链层面运行自动研究循环，跨越多个由仓库衍生且由验证器驱动的环境，仅保留通过筛选的机制。
Four mechanisms survived:
四种机制幸存下来：
> Action Fusion changes how actions execute > Online Context Compact handles compaction during a run > ObservationPack reshapes observation handling > Evidence-Preserving Reducer covers delegated reading
> Action Fusion 改变了动作的执行方式 > Online Context Compact 处理运行期间的上下文压缩 > ObservationPack 重塑了观测处理方式 > Evidence-Preserving Reducer 覆盖了委托读取任务
On the 51-task EdgeBench evaluation, the savings translate to about a third off API cost. In dollars that is an estimated $8.75 to $13.50 per hour against native Codex and Claude Code harnesses, and $4.36 to $5.71 against the baseline harness.
在包含 51 项任务的 EdgeBench 评估中，节省的成本转化为 API 费用降低约三分之一。具体而言，与原生 Codex 和 Claude Code 工具链相比，每小时可节省约 8.75 至 13.50 美元；与基线工具链相比，每小时可节省 4.36 至 5.71 美元。
Because the search runs across many environments rather than one, the retained mechanisms keep working outside the setting that produced them. Code is on GitHub under NVlabs.
由于搜索过程跨越多个环境而非单一环境，所保留的机制在产生它们的环境之外也能持续有效。代码已托管在 GitHub 上的 NVlabs 下。
Paper: https://arxiv.org/abs/2609.20519
论文：https://arxiv.org/abs/2609.20519
Chat with Paper: https://academy.dair.ai/papers/sol-pi-recursively-scaling-auto-research-loops-for-efficient-agent-harness-2609.20519
与论文对话：https://academy.dair.ai/papers/sol-pi-recursively-scaling-auto-research-loops-for-efficient-agent-harness-2609.20519
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力