如何构建长周期自主AI Agent：Basis创始人的实战经验How to Build Long-Horizon Autonomous AI Agents: Lessons from Basis
The Takeaway：可靠的长周期AI Agent成功关键不在于单纯依赖结果验证，而在于编码人类式的流程与行为规范，因为有限评估上的完美结果无法保证真实世界的泛化。
Basis联合创始人Mitch Troyanovsky（该公司是独角兽，专注为会计构建能端到端完成报税等复杂工作的自主Agent）分享了构建能连贯运行数小时甚至数天的Agent的核心方法。他将LLM的局限比作电影《记忆碎片》：拥有巨大工作记忆却几乎没有短期或长期记忆，因此需要harness、笔记和子Agent来维持状态。编码Agent之所以先突破，是因为运行时可验证性以及丰富的训练数据，但非编码领域缺乏这些条件。Basis使用“behavior specs”——用markdown定义期望行为（例如税务研究必须引用原始来源），既作为人类对齐工具，也作为judge的评分标准。“你会看到有人为代码文件抽象不当而抓狂，却对上下文一塌糊涂。English更珍贵，因为它直接影响性能。”Ontology和规范文档对Agent环境至关重要。自我改进将通过行为信号闭环更新harness，最终进入模型权重。技术护城河是暂时的，深度业务嵌入才是长期优势。
The Takeaway: Reliable long-horizon AI agents succeed by encoding human-like processes and behaviors rather than relying solely on outcome verification, because perfect outcomes on limited evals do not guarantee real-world generalization.
Mitch Troyanovsky, cofounder of Basis (a unicorn building autonomous agents for end-to-end accounting work like tax returns), explains how to build agents that stay coherent for hours or days. He compares LLM limitations to the movie Memento: large working memory but no short or long-term memory, so agents need harnesses, notes, and sub-agents to maintain state. Coding agents advanced first due to runtime verifiability and rich training data, but non-coding domains lack easy verification and data. Basis uses "behavior specs" – markdown files defining desired agent behaviors (e.g., always cite primary sources for tax research) that serve as both human alignment tools and judge rubrics. "You'll see people freaking out over a code file that isn't abstracted properly, and yet their context is total shit. The English is more precious because the English affects the performance." Ontologies and canonical documentation are critical for agent environments. Self-improvement will come from closing the loop on signals from behaviors into harness updates, and eventually model weights. Technical moats are temporary; business embedding wins.
Basis联合创始人Mitch Troyanovsky（该公司是独角兽，专注为会计构建能端到端完成报税等复杂工作的自主Agent）分享了构建能连贯运行数小时甚至数天的Agent的核心方法。他将LLM的局限比作电影《记忆碎片》：拥有巨大工作记忆却几乎没有短期或长期记忆，因此需要harness、笔记和子Agent来维持状态。编码Agent之所以先突破，是因为运行时可验证性以及丰富的训练数据，但非编码领域缺乏这些条件。Basis使用“behavior specs”——用markdown定义期望行为（例如税务研究必须引用原始来源），既作为人类对齐工具，也作为judge的评分标准。“你会看到有人为代码文件抽象不当而抓狂，却对上下文一塌糊涂。English更珍贵，因为它直接影响性能。”Ontology和规范文档对Agent环境至关重要。自我改进将通过行为信号闭环更新harness，最终进入模型权重。技术护城河是暂时的，深度业务嵌入才是长期优势。
The Takeaway: Reliable long-horizon AI agents succeed by encoding human-like processes and behaviors rather than relying solely on outcome verification, because perfect outcomes on limited evals do not guarantee real-world generalization.
Mitch Troyanovsky, cofounder of Basis (a unicorn building autonomous agents for end-to-end accounting work like tax returns), explains how to build agents that stay coherent for hours or days. He compares LLM limitations to the movie Memento: large working memory but no short or long-term memory, so agents need harnesses, notes, and sub-agents to maintain state. Coding agents advanced first due to runtime verifiability and rich training data, but non-coding domains lack easy verification and data. Basis uses "behavior specs" – markdown files defining desired agent behaviors (e.g., always cite primary sources for tax research) that serve as both human alignment tools and judge rubrics. "You'll see people freaking out over a code file that isn't abstracted properly, and yet their context is total shit. The English is more precious because the English affects the performance." Ontologies and canonical documentation are critical for agent environments. Self-improvement will come from closing the loop on signals from behaviors into harness updates, and eventually model weights. Technical moats are temporary; business embedding wins.
The Takeaway: Reliable long-horizon AI agents succeed by encoding human-like processes and behaviors rather than relying solely on outcome verification, because perfect outcomes on limited evals do not guarantee real-world generalization.
Mitch Troyanovsky, cofounder of Basis (a unicorn building autonomous agents for end-to-end accounting work like tax returns), explains how to build agents that stay coherent for hours or days. He compares LLM limitations to the movie Memento: large working memory but no short or long-term memory, so agents need harnesses, notes, and sub-agents to maintain state. Coding agents advanced first due to runtime verifiability and rich training data, but non-coding domains lack easy verification and data. Basis uses "behavior specs" – markdown files defining desired agent behaviors (e.g., always cite primary sources for tax research) that serve as both human alignment tools and judge rubrics. "You'll see people freaking out over a code file that isn't abstracted properly, and yet their context is total shit. The English is more precious because the English affects the performance." Ontologies and canonical documentation are critical for agent environments. Self-improvement will come from closing the loop on signals from behaviors into harness updates, and eventually model weights. Technical moats are temporary; business embedding wins.
The Takeaway：可靠的长周期AI Agent成功关键不在于单纯依赖结果验证，而在于编码人类式的流程与行为规范，因为有限评估上的完美结果无法保证真实世界的泛化。
Basis联合创始人Mitch Troyanovsky（该公司是独角兽，专注为会计构建能端到端完成报税等复杂工作的自主Agent）分享了构建能连贯运行数小时甚至数天的Agent的核心方法。他将LLM的局限比作电影《记忆碎片》：拥有巨大工作记忆却几乎没有短期或长期记忆，因此需要harness、笔记和子Agent来维持状态。编码Agent之所以先突破，是因为运行时可验证性以及丰富的训练数据，但非编码领域缺乏这些条件。Basis使用“behavior specs”——用markdown定义期望行为（例如税务研究必须引用原始来源），既作为人类对齐工具，也作为judge的评分标准。“你会看到有人为代码文件抽象不当而抓狂，却对上下文一塌糊涂。English更珍贵，因为它直接影响性能。”Ontology和规范文档对Agent环境至关重要。自我改进将通过行为信号闭环更新harness，最终进入模型权重。技术护城河是暂时的，深度业务嵌入才是长期优势。
查看原文 →
Mitch Troyanovsky, cofounder of Basis (a unicorn building autonomous agents for end-to-end accounting work like tax returns), explains how to build agents that stay coherent for hours or days. He compares LLM limitations to the movie Memento: large working memory but no short or long-term memory, so agents need harnesses, notes, and sub-agents to maintain state. Coding agents advanced first due to runtime verifiability and rich training data, but non-coding domains lack easy verification and data. Basis uses "behavior specs" – markdown files defining desired agent behaviors (e.g., always cite primary sources for tax research) that serve as both human alignment tools and judge rubrics. "You'll see people freaking out over a code file that isn't abstracted properly, and yet their context is total shit. The English is more precious because the English affects the performance." Ontologies and canonical documentation are critical for agent environments. Self-improvement will come from closing the loop on signals from behaviors into harness updates, and eventually model weights. Technical moats are temporary; business embedding wins.
The Takeaway：可靠的长周期AI Agent成功关键不在于单纯依赖结果验证，而在于编码人类式的流程与行为规范，因为有限评估上的完美结果无法保证真实世界的泛化。
Basis联合创始人Mitch Troyanovsky（该公司是独角兽，专注为会计构建能端到端完成报税等复杂工作的自主Agent）分享了构建能连贯运行数小时甚至数天的Agent的核心方法。他将LLM的局限比作电影《记忆碎片》：拥有巨大工作记忆却几乎没有短期或长期记忆，因此需要harness、笔记和子Agent来维持状态。编码Agent之所以先突破，是因为运行时可验证性以及丰富的训练数据，但非编码领域缺乏这些条件。Basis使用“behavior specs”——用markdown定义期望行为（例如税务研究必须引用原始来源），既作为人类对齐工具，也作为judge的评分标准。“你会看到有人为代码文件抽象不当而抓狂，却对上下文一塌糊涂。English更珍贵，因为它直接影响性能。”Ontology和规范文档对Agent环境至关重要。自我改进将通过行为信号闭环更新harness，最终进入模型权重。技术护城河是暂时的，深度业务嵌入才是长期优势。