精选889月4日 08:00Cursor 博客（web_list）技巧与观点

Basis 如何用 Cursor 构建长周期会计 Agent

How Basis builds long-horizon accounting agents with CursorSep 4, 2026

原文

发到 X

推荐理由

Agent 长周期工作的过程控制是行业痛点，这篇分享的行为规范方法与 Cursor 实操细节极具参考价值，建议做 Agent 的同学重点阅读。

Blog / customers

博客 / 客户案例

Contact sales

联系销售

Basis builds AI agents specifically for accountants. They autonomously complete long-horizon, complex accounting workflows in the background and return review-ready outputs, so accounting teams can focus on judgment and client service.

Basis 专为会计师构建 AI 智能体。它们能在后台自主完成长周期、复杂的会计工作流，并返回可供审核的输出结果，从而使会计团队能够专注于专业判断和客户服务。

These agents take on multi-hour, high-stakes work for leading accounting teams: month-end close, corporate and partnership tax returns, audit planning and fieldwork. Basis built the company on Cursor from day one. They treat the context the agent reads (prompts, skills, instructions, tool descriptions) with the same rigor as code, and Cursor is where they read and revise it.

这些智能体为顶尖会计团队承担耗时数小时且高风险的工作：月末结账、企业与合伙企业税务申报、审计规划及外勤工作。Basis 从成立之初就基于 Cursor 构建公司。他们将智能体读取的上下文（提示词、技能、指令、工具描述）视为与代码同等重要的内容，而 Cursor 正是他们阅读和修订这些内容的地方。

#Work that cannot be reduced to a single prompt

#无法简化为单一提示词的工作

Long-horizon does not just mean the agent runs for several hours. It means hundreds of decisions across a trajectory, with later steps often depending on earlier ones. The system must preserve relevant state, incorporate the results of tool calls, and recover from failures, sometimes across more information than fits in a single context window. Errors can compound. An early mistake can affect later research, calculations, tool calls, and artifacts, while the final result may not reveal where the problem began.

长周期不仅仅意味着智能体运行数小时。它意味着在整个执行轨迹中做出数百次决策，且后续步骤往往依赖于前面的步骤。系统必须保留相关状态，整合工具调用的结果，并从失败中恢复，有时涉及的信息量甚至超过单个上下文窗口的容量。错误可能会累积放大。早期的失误可能影响后续的研究、计算、工具调用和产物生成，而最终结果可能无法揭示问题究竟出在哪里。

Accounting makes this harder for three reasons:

会计领域使这一挑战更加复杂，原因有三：

Many outcomes do not have a cheap, objective test.
Ground-truth examples drawn from real production work are expensive to create and difficult to scale.
A final outcome may take hours or days to produce and review.

许多结果没有廉价且客观的测试方法。
源自真实生产工作的基准示例创建成本高且难以规模化。
最终结果的产出和审核可能需要数小时甚至数天。

Even a correct final result can hide an unreliable process. An agent might reach the correct tax return without research authority, extract the correct number without preserving its source, or produce a usable workbook through a process that will not generalize. Outcome evaluation still matters, but it is expensive to run, and it cannot explain every consequential decision inside a long trajectory.

即使最终结果是正确的，也可能掩盖了不可靠的过程。智能体可能在缺乏研究依据的情况下得出正确的税务申报表，或在未保留来源的情况下提取正确的数字，或通过无法泛化的过程生成可用的工作簿。结果评估仍然重要，但运行成本高昂，且无法解释长轨迹中每一个关键决策背后的原因。

#Context is a production input

#上下文是生产级输入

The agent's final output is only one part of the system. Its behavior depends on the context it receives throughout the work: instructions, domain knowledge, examples, tool descriptions, skills, memory, and other runtime information. That context is written in natural language, so engineers have to read it.

智能体的最终输出只是系统的一部分。其行为取决于在工作过程中接收到的上下文：指令、领域知识、示例、工具描述、技能、记忆以及其他运行时信息。这些上下文以自然语言编写，因此工程师必须阅读它们。

A traditional program interprets the same valid code the same way regardless of how neatly the files are organized. With language models, the organization and wording of context change what the model does next. A vague sentence, a buried exception, or a misleading example can change production behavior. Generating a context file and shipping it without reading it is a production risk.

传统程序对相同的有效代码始终以相同的方式解释，无论文件组织得多么整齐。而对于语言模型，上下文的组织和措辞会改变模型的后续行为。一句模糊的话、一个隐蔽的异常或一个误导性的示例都可能改变生产环境中的行为。生成上下文文件并在未阅读的情况下直接部署，会带来生产风险。

#Behavior specs make the standard explicit

#行为规范使标准显式化

A behavior spec is a Markdown file that defines recurring conduct expected from an agent in a specific situation. It is written for the people and judges reviewing a recorded trajectory. It is not a prompt, and it is not shown to the agent.

行为规范是一个 Markdown 文件，定义了智能体在特定情境下应遵循的重复性行为。它是为审阅录制轨迹的人员和评估者编写的。它不是提示词（prompt），也不会展示给智能体。

A useful spec makes clear when the behavior applies, what evidence the agent should inspect, what decision it should make, what action should follow, what to do when evidence is incomplete, and what failure looks like. The goal is to make the behavior judgeable without scripting every step.

一份有用的规范应明确说明：该行为适用的场景、智能体应检查的证据、应做出的决策、随后的行动、证据不完整时的处理方式，以及失败的表现形式。其目标是使行为可被评估，而无需将每个步骤都脚本化。

The judge receives the spec, the observable trajectory, and the evidence (tool calls, artifacts, retrieved sources, decision records). It returns true, false, or NA. That lets the team evaluate selected parts of the process without a complete ground-truth answer for the whole task.

评估者接收规范、可观察的轨迹以及证据（工具调用、产物、检索到的来源、决策记录）。它返回 true、false 或 NA。这使得团队能够在没有整个任务完整真实答案的情况下，评估流程的选定部分。

#Cursor is where they revise the agent

#Cursor 是修订智能体的地方

Basis uses Cursor to build and refine its agents. An engineer has a behavior spec open in Markdown. They inspect a sentence, ask a model whether it is too vague or too brittle, revise the passage, and preview the finished document in the same window. They use the same environment to refine the prompts and context the agent actually sees: skills, instructions, tool descriptions.

Basis 使用 Cursor 来构建和优化其智能体。工程师在 Markdown 中打开行为规范。他们检查某句话，询问模型它是否过于模糊或脆弱，修订段落，并在同一窗口中预览完成的文档。他们使用相同的环境来优化智能体实际看到的提示词和上下文：技能、指令、工具描述。

What makes Cursor the place for that work:

使 Cursor 成为该工作场所的因素包括：

A real editor, so one can read and revise the wording of context and specs.
Markdown preview (edit and live preview together). Mitch Troyanovsky, cofounder of Basis, called this an underrated differentiator for iterating specs, skills, and other Markdown docs.
Working with a model directly, in the same environment as the text.
Easy model switching while you iterate.
Side-by-side: the file and the agent window as a loop. Everyone has an agent window. The difference is being able to inspect and change the context.

一个真正的编辑器，允许人们阅读并修订上下文和规范的措辞。
Markdown 预览（编辑与实时预览结合）。Basis 联合创始人 Mitch Troyanovsky 称这是迭代规范、技能和其他 Markdown 文档时被低估的差异点。
在与文本相同的环境中直接与模型协作。
在迭代过程中轻松切换模型。
并列显示：文件和智能体窗口形成闭环。每个人都有智能体窗口。区别在于能够检查和更改上下文。

Cursor is where we inspect the context that shapes the agent, and revise it until the behavior holds.

Cursor 是我们检查塑造智能体的上下文的地方，并对其进行修订，直到行为符合要求。

Mitch Troyanovsky

Co-founder of Basis

Basis 联合创始人

#The development loop

#开发循环

Basis engineers write and refine behavior specs in Cursor. The agent runs in the Basis runtime, and a judge evaluates the recorded trajectory against the spec.

Basis 工程师在 Cursor 中编写并完善行为规范。代理在 Basis 运行时环境中运行，裁判根据规范评估记录的轨迹。

The team agrees on a recurring behavior worth measuring.
An engineer writes or refines the behavior spec in Cursor.
The agent performs its work in production, producing a recorded trajectory.
A judge evaluates each behavior against the spec, returning true, false, or NA.
A false verdict identifies a gap between the intended behavior and the runtime implementation.
The team updates the runtime context, tools, prompts, or execution framework. That wording gets revised in Cursor.
The team runs the agent again and measures whether the behavior improves.

团队就一项值得衡量的重复性行为达成一致。
工程师在 Cursor 中编写或完善行为规范。
代理在生产环境中执行工作，生成记录的轨迹。
裁判根据规范评估每项行为，返回 true、false 或 NA。
错误的判定揭示了预期行为与运行时实现之间的差距。
团队更新运行时上下文、工具、提示词或执行框架。这些措辞会在 Cursor 中进行修订。
团队再次运行代理，并衡量行为是否有所改善。

The spec and the runtime stay separate. The spec is the standard. The implementation changes until the agent meets it consistently.

规范与运行时保持分离。规范是标准。实现方式会不断调整，直到代理能够持续符合规范。

The behavior-spec approach grew from Basis's experience building production agents for accounting. Basis and Braintrust released it as an open standard so other teams can define and evaluate agent behavior using the same general format.

行为规范方法源于 Basis 构建会计领域生产级代理的经验。Basis 和 Braintrust 将其作为开放标准发布，以便其他团队使用相同的通用格式来定义和评估代理行为。

A correct tax answer can still hide a bad process. I want to know if the agent checked primary authority, not just if the return is right. The spec is how we judge that.

正确的税务答案仍可能掩盖糟糕的处理过程。我想知道代理是否查阅了首要权威来源，而不仅仅是申报表是否正确。规范是我们据此进行评判的方式。

Mitch Troyanovsky

Co-founder of Basis

Basis 联合创始人

#The work itself is the proof

#工作本身即是证明

This is what that work looks like in production.

这就是该工作在真实生产环境中的样子。

Basis agents perform 5+ hours of work on a single deliverable.
On a Form 1065 partnership return, work that can take roughly 30 to 40 hours of human time can be completed by a Basis agent in roughly 6 to 7 hours.
Basis is trusted by 40% of the top 25 firms, and by leading accounting firms more broadly.

Basis 代理对单个交付物执行超过 5 小时的工作。
在处理 Form 1065 合伙企业申报表时，原本需要人类大约 30 到 40 小时的工作，Basis 代理仅需大约 6 到 7 小时即可完成。
Basis 受到前 25 大事务所中 40% 的信任，并在更广泛的领先会计师事务所中获得认可。

The strongest proof is the work itself: agents making many decisions over long trajectories, delivering work professional accountants review and use.

最有力的证明就是工作本身：代理在长轨迹中做出大量决策，交付专业会计师审查和使用的工作成果。

As agents take on longer and more consequential work, their context becomes a production input. Engineers have to inspect it, understand it, and revise it.

随着代理承担更长且更具影响力的工作，其上下文成为生产输入。工程师必须检查、理解并修订它。

Cursor is where Basis maintains that context. Behavior specs make selected expectations explicit. Braintrust evaluates whether those behaviors appeared in real trajectories. Failures tell the team what to change in the runtime.

Cursor 是 Basis 维护该上下文的地方。行为规范使选定的期望显式化。Braintrust 评估这些行为是否出现在真实轨迹中。失败告诉团队需要在运行时做出哪些更改。