Cresta 如何借助 Claude Agent SDK 将客户体验专长转化为 Agent 构建器
英文标题：How Cresta turned CX expertise into an agent builder on the Claude Agent SDK
企业级客户体验 AI 平台 Cresta 基于 Claude Agent SDK 打造了自然语言 Agent 构建器 Conductor，帮助团队将复杂业务上下文转化为高可靠的生产级 Agent。早期实践表明，Conductor 将初始部署时间缩短了约 50%。该系统充当“元 Agent（meta-agent）”，不仅平衡了关键业务流的确定性与对话灵活性，还通过端到端评测闭环与经验资产沉淀加速开发。这为开发者展示了如何将通用 Agent 框架与垂直行业专长深度融合，以应对 Agent 上线后长期迭代的工程挑战。
15.7 KB
204.6 KB
228.5 KB
184.1 KB
176.1 KB
403 B
72.5 KB
中文
在我们的系列专题**《初创团队如何基于 Claude 构建产品》**（How startups build with Claude）中，我们深入探讨了 AI 产品背后的架构与工程决策。本文将剖析 Cresta 如何将其在打造原生 AI 客户体验（CX）方面的专长，沉淀为名为 Conductor 的 Agent，进而帮助开发者构建和优化其他 Agent。
速览
名称 | Cresta 成立时间 | 2017 年 CEO | Ping Wu 业务增长 | 累计融资超 2.7 亿美元；年度经常性收入（ARR）超 1 亿美元；助力联合航空（United Airlines）、CVS Health 及万豪国际（Marriott）等企业提供卓越的客户体验
构建用于客户体验场景的企业级 AI Agent，需要深刻理解跨各种客户交互场景的深层上下文。例如，客户可能会：
- 请求退款却未说明具体是哪一笔订单；
- 在同一条消息中同时报告账单问题和登录故障；或者
- 在退货窗口期关闭后才提出退款请求。
为了做出正确响应，Agent 可能需要核对客户购买记录、退款政策、账户历史或支付状态。
Cresta 每天都在海量交互中处理这类场景。其平台为客户体验提供动力，驱动能够自主处理会话的 AI Agent，为协同工作的人工坐席提供实时指引，并沉淀对话智能（conversation intelligence）以向企业指出需要改进的环节。
构建和维护客户体验 Agent 需要企业在运营需求上做出关键决策：Agent 需要了解哪些信息？必须访问哪些系统？其行为在何处应当保持灵活，在何处必须具备确定性？Cresta 希望让更多团队能够直接运用这些判断标准，而无需各自从零起步。
Conductor 应运而生：这是一个自然语言 Agent 构建器，可帮助团队将复杂的业务上下文转化为可直接用于生产环境的 Agent。用户只需描述他们想要构建的内容，Conductor 就会引导整个开发流程——从可靠的基础蓝图一直到实现、评估和持续优化。
Cresta 最初使用 Claude Sonnet 和 Claude Opus 构建了 Conductor 的早期内部版本，用于帮助内部团队支持客户部署。当时未解决的开放性问题在于：Cresta 能否将该系统打造成一款成熟产品，供客户直接在平台内自行构建 Agent？
为了执行开放式的开发任务，Conductor 将 Claude Agent SDK 用作通用的执行框架（harness），以此收集上下文、调用工具、编写并运行代码，并根据执行结果动态调整。
“我们在内部将 Claude Code 用作开发工具，因为它在构建 Agent 的软件工程方面非常高效，”Cresta Conductor 工程主管 Renjie Li 表示，“Claude Agent SDK 为我们提供了一种以编程方式利用该 Agent 研发框架的途径。借助 Conductor，我们将其与 Cresta 深厚的对话智能、客户体验专长、评测体系和运行时环境连接起来，为用户提供我们从第一天起就梦寐以求的 Agent 构建体验。”
Cresta 指出，在 Cresta 及其合作伙伴的早期落地案例中，Conductor 将初始部署时间缩短了大约一半。
帮助构建下一个 Agent 的“元 Agent”
Conductor 是一个基于已验证的客户体验 Agent 最佳开发实践构建的、具备自进化能力的“元 Agent”（meta-agent）。随着团队使用它构建和打磨越来越多的 Agent，他们能够捕获自身经过验证的模式、工作流以及专家判断。这形成了一个知识飞轮，使未来的开发更加敏捷和一致。
上线之后，生产环境中的实际交互、工作流执行结果以及用户反馈，都会转化为推动下一次迭代改进的新信号。
Conductor 可帮助开发者将在构建 Agent 过程中沉淀的经验转化为可复用的“记忆资产”（memory artifact），封装其中的交互模式、业务规则和评估方法。随后，他们可以将这些成果作为技能（skill）共享给团队其他成员，从而让其他人基于同样验证过的最佳实践进行构建。
让关键决策变得更加轻松
编写最初的提示词（Prompt）只是构建实用 Agent 的冰山一角。更困难的决策之一在于：对话在何处应保持开放灵活，而在何处必须受限于明确的业务规则和严格受控的工具行为。
“你不能走向极端的确定性，否则你会陷入试图穷举每一种可能分支的庞大决策树……但面对对风险容忍度极低的强监管行业，某些对企业至关重要的工作流必须确保万无一失，”Renjie 表示，“Conductor 有助于梳理并固化这些确定性要求，同时提供足够的灵活性来优化整体体验。”
Conductor 借助历史对话数据，协助团队明确人工应当介入的边界，以及模型可以自主发挥适应的空间。一旦界定了这些边界，开发者仍需要针对既定要求对确定性模块进行测试，并在底层业务规则发生变更时及时修订这些测试用例。
Cresta 如何评估 Conductor
Conductor 的衡量标准取决于其构建 Agent 的水平。为此，Cresta 会让 Conductor 执行一系列具体的构建任务：
- 构建一个全新 Agent
- 为现有 Agent 编写测试用例
- 修改已存在的 Agent
- 进行根本原因分析（Root Cause Analysis）
顺利完成任务并非唯一的达标要求。团队还会考量执行结果、Conductor 达成结果的路径，以及它在整个过程中消耗的资源。
| 评估维度 | 回答的问题 |
|---|---|
| 执行结果（Outcome） | Conductor 是否生成了符合要求且切实可用的成果物？ |
| 执行路径（Execution path） | 它是否调用了合适的工具，并获取了必要的上下文？ |
| 交付质量（Quality） | 结果与精选的基准参考相比表现如何？ |
| 资源消耗（Resource use） | 该任务消耗了多少时间和模型调用量？ |
这些基准测试任务同样扮演着安全网的角色。每当新的 Claude 模型发布或 Conductor 框架升级时，Cresta 都会重新运行这套评测。团队使用完全相同的任务集和评分体系，在代码合并前精准识别哪些能力得到了提升，哪些地方仍需改进。
Conductor 本身也内置了这套严格的评测规范，让终端用户能够检测自己的 Agent 是否按预期工作。Conductor 会将构建阶段捕捉到的业务需求和边缘情况，自动转化为评估模块中的测试用例。
这样一来，随着业务政策、对接系统和对话模式的演变，团队可以持续监控那些“绝不能出错”的工作流。这为他们在业务变迁时更新 Agent 提供了一种切实可行的方法，而不会破坏已固化工作流的可靠性。
基于 Claude Agent SDK 构建 Conductor
Cresta 将 Conductor 设计为专用于客户体验（CX）领域的 Agent 开发控制平面（control plane）。它汇集了对话数据、领域专长和工程开发实践，并通过策略控制、可观测性、验证和反馈驱动的持续改进来管理 Agent 的每次运行。
Claude Agent SDK 作为通用执行框架构成了 Conductor 的底层支撑。它负责处理收集上下文、调用工具、编写并运行代码等通用任务；Conductor 则在此之上叠加了 Cresta 的客户体验工作流、专有工具和领域上下文。
团队针对 Conductor 需要执行的任务对 SDK 进行了严格评估，包括多步骤工作流、庞大上下文处理以及高频工具调用。审查范围覆盖了数据隐私、多租户架构、组织级密钥分发、权限控制以及整个实现链路的可观测性。
“Anthropic 为我们开展 Agent 化软件开发提供了强大而通用的基础底座。这使我们能够将工程研发资源集中倾斜到 Cresta 能为客户创造最持久价值的领域，”Cresta 工程副总裁 Xiangru Chen 说道。
将 Cresta 的工程专长直接交付给客户
Conductor 最初源自 Cresta 内部，用于帮助其前沿部署团队（forward-deployed team）明确哪些开发任务可以标准化复用、上下文在哪些环节最为关键，以及构建者在整个流程中需要审视哪些指标。该工具在内部任务中的成功，让团队充满信心地将同样的能力推向客户与合作伙伴。
对于任何遇到“客户定制 Agent 的交付需求速度快于工程师手工开发速度”的团队而言，Cresta 的实践都是极具价值的参考范例：无论面对普通 Agent 还是元 Agent，最初构建完成之后的工作往往才是最艰难的部分。
了解 Conductor 如何将业务上下文转化为更优的 Agent 设计、实现与持续改进。
来自 Cresta 团队的最佳实践
坚持吃自己的狗粮（Dogfooding） | Conductor 最初只是内部工具，直到其产出水准达到了具备深厚 AI 专长工程师的代码水准后，才被正式包装为对外产品。 聚焦自身的独特壁垒 | Cresta 划定了一条清晰的界线：只专注提供唯有他们才能带来的核心壁垒（即独有的对话数据与对客户体验的深刻洞察），直接采用 Agent SDK 提供的现成能力，绝不在非核心优势领域重复造轮子。 明确 Agent 中哪些部分必须保持确定性 | 将绝不能出错的企业关键业务工作流严格固化，其余部分保持灵活；否则，你将无法获得期望的交互式 Agent 体验。 推向生产环境只是最容易的一步 | 最初的构建工作仅占整体工作量的 20% 左右，其余 80% 都在于随后的测试、优化和持续迭代；因此，在设计工具链时，应围绕迭代闭环而非仅仅围绕上线发布展开。
English / 原文
In our series,How startups build with Claude , we look at the decisions behind AI products. Here, we explore how Cresta brought its expertise building AI-native customer experience into Conductor , an agent that helps developers build and improve other agents.
The quick pitch
Name | Cresta
Founded | 2017
CEO | Ping Wu
Growth | $270M+ raised. $100M+ ARR. Helping organizations like United Airlines, CVS Health, and Marriott deliver exceptional customer experiences
Building enterprise AI agents for customer experience use cases requires understanding the context embedded across a wide range of customer interactions. For example, customers may:
- Ask for a refund without saying which purchase they mean,
- Report a billing problem and a login issue in the same message, or
- Request a refund after the return window has closed.
To respond correctly, the agent may need to check what the customer bought, the refund policy, their account history, or the payment status.
Cresta sees these interactions at volume. Its platform powers customer experience, AI agents that handle conversations on their own, provide real-time guidance for the human agents working alongside them, and surface conversation intelligence that shows the business where to improve.
Building and maintaining customer experience agents requires organizations to make key decisions on operating requirements: what the agent needs to know, which systems it must access, and where its behavior should be flexible or deterministic. Cresta wanted to make that judgment available to more teams without each one starting from scratch.
Conductor is the result: a natural language agent builder that helps teams turn complex business context into production-ready agents. Users describe what they want to build, and Conductor guides the work from a grounded blueprint through implementation, evaluation, and optimization.
Cresta originally built an early version of Conductor using Claude Sonnet and Claude Opus to help internal teams support customer deployments. The open question was whether Cresta could turn that system into a product customers could use to build agents directly within the platform.
To carry out open-ended development work, Conductor uses the Claude Agent SDK as a general-purpose harness to gather context, use tools, write and run code, and adapt based on the results.
“We used Claude Code internally as a dev tool because it was very effective at the software development aspects of building an agent,” said Renjie Li, Engineering Lead for Cresta Conductor. “The Claude Agent SDK gave us a way to use that same agentic development harness programmatically. With Conductor, we’ve connected it with Cresta’s deep conversation intelligence, CX expertise, evaluations, and runtime to give users the agent-building experience we wish we’d had from day one.”
Cresta reports that in early use cases across Cresta and its partners, Conductor cut initial deployment time roughly in half.
A meta-agent that helps build the next agent
Conductor is a self-improving meta-agent built on established best practices for CX agent development. As teams build and refine more agents with it, they capture their own proven patterns, workflows, and expert judgment. This creates a knowledge flywheel that makes future development faster and more consistent.
After launch, production interactions, workflow outcomes, and feedback become new signals for the next improvement.
Conductor helps builders turn what they learned from an agent build into a reusable memory artifact, capturing patterns, business rules, and evaluation approaches. They can then share it as skill with the rest of the team, so others can build from the same proven practices.
Making the important decisions easier
Writing the first prompt is only one part of building a useful agent. One of the harder calls is where a conversation should stay flexible and where it needs explicit business rules and controlled tool behavior.
"You don't want to go too deterministic, otherwise you create a giant decision tree trying to map out every possible branch… but there are enterprise-critical workflows that you need to make sure work all the time, especially for highly regulated industries with low risk tolerance," said Renjie. "Conductor helps discover and harden those deterministic requirements and the implementation with the flexibility to improve the overall experience."
Conductor draws on historical conversations to help teams identify where a human should stay in the loop and where the model can be given room to adapt. Once those boundaries are set, developers still need to test the deterministic pieces against the requirements they're meant to enforce, and revise those tests as the underlying business rules change.
How Cresta evaluates Conductor
Conductor is measured on how well it builds agents. To that end, Cresta runs Conductor through a set of build tasks:
- Build a new agent
- Write a test case for one
- Change an agent that already exists
- Perform root cause analysis
Finishing the task is not the only bar. The team also looks at the outcomes, how Conductor got there, and what it used along the way.
| Evaluation | Question it answers |
|---|---|
| Outcome | Did Conductor produce the requested, usable artifact? |
| Execution path | Did it use the appropriate tools and required context? |
| Quality | How does the result compare with a curated reference? |
| Resource use | How much time and model usage did the task require? |
Those same tasks function as a safety net. When a new Claude model is launched or the Conductor framework is updated, Cresta reruns its evaluations. It uses the same tasks and same scoring to identify what improved and what still needs work before anyone commits to the change.
That same evaluation discipline is built into Conductor, which allows its users to check and see if their agents are performing as intended. Conductor turns requirements and edge cases captured during the build phases into tests within its evaluation module.
Teams can then monitor the workflows that need to work every time as policies, connected systems, and conversation patterns change. That gives them a practical way to update an agent as the business changes without losing the reliability of the workflows they have already hardened.
Building Conductor on the Claude Agent SDK
Cresta designed Conductor as the CX-specific control plane for agent development. It brings together conversation data, domain expertise, and development practices, then governs each agent run through policy, observability, verification, and feedback-driven improvement.
The Claude Agent SDK sits beneath Conductor as a general-purpose execution harness. It manages the work of gathering context, calling tools, and writing and running code; Conductor layers Cresta's CX workflows, tools, and domain context on top.
The team evaluated the SDK against the work Conductor needed to perform: tasks with multiple steps, substantial context, and repeated tool use. Its review covered data privacy, tenant architecture, organization-level key distribution, controls, and observability across the implementation.
"Anthropic gives us a strong, general-purpose foundation for agentic software development. That lets us put our engineering investment where Cresta creates the most durable customer value," said Xiangru Chen, VP of Engineering at Cresta.
Putting Cresta's engineering expertise directly in customers' hands
Conductor started inside Cresta, helping its own forward-deployed team learn which development tasks could become repeatable, where context mattered most, and what builders needed to inspect along the way. Its success in that internal work gave the team confidence to bring the same capabilities to customers and partners.
For any team whose product deploys custom agents faster than its engineers can hand-build them, Cresta's approach is a useful reference point: with agents or meta-agents, the work after the initial build is often the hardest part.
See how Conductor turns business context into better agent design, implementation, and ongoing improvement.
Best practices from the Cresta team
Dogfood your products | Conductor ran as an internal tool first, and only became a product once its output matched what an engineer with deep AI expertise would build.
Focus on your unique advantages | Cresta drew a hard line between the expertise only they could bring (their conversation data and understanding of customer experience) and what they could take off the shelf from the Agent SDK, and refused to build anything on the wrong side of it.
What parts of your agent need to be deterministic | Harden the enterprise-critical workflows that can never go wrong and leave the rest flexible, or you won't end up with the interactive agentic experience you want.
Getting to production is the easy part | The initial build is only about 20% of the effort, the other 80% is the testing, optimization, and continuous improvement that follows, so design your tooling around the iteration loop, not the launch.