让 Agent 的执行轨迹，成为模型持续进化的训练信号
NeoHorse-1 是面向文本 Agent harness 的开源模型家族，也是通往递归自我改进的初始原型。它通过路由引导的课程学习与在线策略蒸馏，让工具调用、代码与任务执行结果回流到后训练过程。
基元律动
tokenrhythm.ai
All Models, One Mind
OpenSquilla 任务场景
给一个主题，它自己去找、去核、去比
从一堆论文里，捞出真正有用的结论
文件再多，也能自己归类、提取、整理
读懂项目，再把任务落到代码里
模型即服务，让 Agent 用上多模型
少做接入和维护，把时间留给产品。模型选择、协作与用量管理，由一套接口完成。
开始使用-
一次接入主流模型
不用逐家申请、充值和维护。
-
每次都用对模型
按任务需要，选择真正合适的能力。
-
复杂任务自动协作
路由与多模型融合，由系统自动完成。
-
每一笔成本都清楚
花在哪个模型、用了多少 Token，一目了然。
研究与进展
我们把路由、评测与 harness 设计的结论公开发表 —— 产品里跑的，就是论文里写的那一套。
01
NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness
2026
Agent Systems
阅读全文
NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness
NeoHorse-1 combines agentic post-training with routing-derived curricula and capability-guided data allocation to close an evaluation-selection-update loop toward recursive self-improvement.
阅读全文
02
OpenSquilla: Token-Efficient Agent = Models + Routing Harness
2026
Agent Systems
阅读全文
OpenSquilla: Token-Efficient Agent = Models + Routing Harness
OpenSquilla combines step-level singleton routing and multi-model ensemble routing in a learnable agent harness to preserve or improve task quality while substantially reducing execution cost.
阅读全文
03
Agentic Routing: The Harness-Native Data Flywheel
2026
Agent Systems
阅读全文
Agentic Routing: The Harness-Native Data Flywheel
Agentic routing turns execution traces and task outcomes into a harness-native data flywheel for continuously improving model selection, orchestration, and efficiency.
阅读全文
04
From Question Answering to Task Completion: A Survey on Agent System and Harness Design
2026
Agent Systems
阅读全文
From Question Answering to Task Completion: A Survey on Agent System and Harness Design
A systematic survey of agent systems and harness design, reframing progress from answering questions toward reliably completing long-horizon tasks.
阅读全文
05
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks
2026
Machine Learning
阅读全文
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks
Claw-SWE-Bench evaluates OpenClaw-style agent harnesses on realistic coding tasks while measuring capability, reliability, and the cost of end-to-end execution.
阅读全文
06
Sibyl-AutoResearch: Autonomous Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators
2026
Agent Systems
阅读全文
Sibyl-AutoResearch: Autonomous Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators
Sibyl-AutoResearch turns trial signals, failures, and evidence into auditable updates to later plans, validation, claims, scheduling, memory, and harness behavior.
阅读全文