MIT提出JAZ极简Agent框架:以代码为记忆与自改进机制
Keep your agent harness minimal, folks.
Agent架构师必读,这篇论文用实验证明了极简主义的力量,把记忆和自改进塞进代码循环里比外挂子系统更省钱效果更好。
Keep your agent harness minimal, folks.
各位,请尽量保持智能体框架的简洁。
This paper from MIT CSAIL shows why.
麻省理工学院计算机科学与人工智能实验室(CSAIL)的这篇论文展示了原因。
(bookmark it)
(收藏它)
While being very minimal it looks effective and promising for building self-improving agents.
尽管非常精简,但它在构建自我改进型智能体方面看起来有效且充满希望。
They present JAZ, an agent framework with one primitive, invoke. The LLM writes code, can call invoke recursively, and sees all of its inputs and history as variables in the code environment.
他们提出了 JAZ,一个仅包含一个原语 invoke 的智能体框架。LLM 编写代码,可以递归调用 invoke,并将所有输入和历史记录作为代码环境中的变量来查看。
With prompting only and no memory system, it beats Letta (MemGPT) by 8% at half the cost on the recall-heavy part of StuLife.
仅通过提示词且无需记忆系统,在 StuLife 中侧重回忆的部分,它以一半的成本比 Letta (MemGPT) 高出 8%。
On AppWorld it beats ACE, a self-improvement method, by 4% at lower cost.
在 AppWorld 上,它以更低的成本比 ACE(一种自我改进方法)高出 4%。
Memory and self-improvement are usually built as separate subsystems. Here both are code the agent writes inside the loop, and hooks handle the constraints and monitoring you want to enforce.
记忆和自我改进通常被构建为独立的子系统。在这里,两者都是智能体在循环内编写的代码,而钩子(hooks)则处理你想要执行的约束和监控。
Paper: https://arxiv.org/abs/2609.26891
论文:https://arxiv.org/abs/2609.26891
Explore Paper: https://academy.dair.ai/papers/harness-as-a-language-a-minimalist-agent-framework-with-maximal-expressivity-2609.26891
探索论文:https://academy.dair.ai/papers/harness-as-a-language-a-minimalist-agent-framework-with-maximal-expressivity-2609.26891
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