In this episode of Zero to Agent in 30 Minutes, Jayeeta Putatunda, forward deployed AI engineering lead at Turing, builds a multi-agent workflow that turns a daily flood of financial headlines into a structured analyst briefing. Financial analysts already have deep internal research, coverage assignments, and market views, but keeping that context current as new information arrives every day and surfacing which of it actually deserves an analyst’s attention is a harder problem.在本期《30分钟打造智能体》节目中，Turing 前线部署 AI 工程负责人 Jayeeta Putatunda 构建了一个多智能体工作流，能将每日海量的金融头条新闻转化为结构化的分析师简报。金融分析师通常已有深厚的内部研究、覆盖任务和市场观点，但随着每日新信息的涌入，如何保持这些背景信息的时效性，并筛选出真正值得分析师关注的内容，是一个更具挑战性的难题。
How to build a source-backed briefing agent, step by step如何逐步构建一个基于来源的简报智能体
- Define the scope. Start by setting the analyst’s focus area, research questions, and time window, whether that’s the last one day, seven days, or 30 days. Narrowing the scope up front, including a list of preferred sources, keeps the agent’s web searches directional instead of generic.定义范围。首先设定分析师的关注领域、研究问题以及时间窗口（无论是过去一天、七天还是三十天）。提前缩小范围并列出首选来源，可以使智能体的网络搜索更具针对性，而非泛泛而谈。
- Plan the coverage. A coverage planner agent breaks the research question into discrete sections, such as market backdrop or company catalysts, so the search that follows can run in parallel rather than one long sequential query.规划覆盖范围。覆盖规划智能体将研究问题拆解为独立的板块，如市场背景或公司催化剂，以便后续的搜索能够并行执行，而不是进行冗长的顺序查询。
- Search and gather. A news researcher agent runs multiple queries against the preferred sources first, then falls back to a general search if the preferred sources don’t return enough results.搜索与收集。新闻研究智能体会先针对首选来源进行多次查询，如果首选来源返回的结果不足，则会回退到通用搜索。
- Validate the sources. A validation agent checks each link for a working, clickable URL and a correct publication date, removes duplicate stories covering the same news, and filters out paywalled pages that won’t return usable content.验证来源。验证智能体会检查每个链接是否有效、可点击且发布日期正确，剔除报道相同新闻的重复内容，并过滤掉无法获取可用内容的付费墙页面。
- Generate the briefing. A briefing writer agent assembles the validated developments into a set structure, including an executive summary, key bullet points, and a section-by-section breakdown of what each development means for the analyst and what to watch next.生成简报。简报撰写智能体会将经过验证的进展汇总成固定结构，包括执行摘要、关键要点，以及针对每一项进展对分析师意味着什么、后续应关注什么的逐项分析。
- Capture feedback into memory. A feedback agent logs corrections such as formatting or terminology preferences and saves them to a memory database, so the next briefing run applies those preferences automatically instead of requiring the analyst to re-prompt.将反馈存入记忆。反馈智能体会记录格式或术语偏好等修正意见，并将其保存到记忆数据库中，这样下一次生成简报时会自动应用这些偏好，无需分析师重新提示。
Jayeeta built the entire stack on open source models so newcomers can run it without an API key, and she recommends starting with a smaller model before scaling up. The takeaway extends well beyond finance. Building single-purpose agents rather than one large agent that handles every task means a failure at one stage doesn’t force a restart of the whole pipeline, and each agent’s output stays easier to trace and debug.Jayeeta 使用开源模型构建了整个技术栈，因此新手无需 API 密钥即可运行；她建议在扩展规模前先从小模型开始尝试。这一经验远超金融领域。构建单一用途的智能体而非一个处理所有任务的大型智能体，意味着某个阶段的失败不会导致整个流水线重启，且每个智能体的输出也更易于追踪和调试。
The full code base, including the sample data and the UI shown in the demo, is available in Jayeeta’s GitHub repo, so readers can clone it and run the briefing agent on their own systems.完整的代码库（包括示例数据和演示中展示的 UI）已上传至 Jayeeta 的 GitHub 仓库，读者可以克隆并在自己的系统上运行该简报智能体。
Coming this week本周预告
This week, Maxim Salnikov joins Zero to Agent in 30 Minutes to build a supply chain for agent context. He’ll show how to source approved packages from a trusted registry, pin and hash-verify them on any harness, and enforce org policy with a CI gate that can’t be bypassed.本周，Maxim Salnikov 将加入《30分钟打造智能体》，共同构建智能体上下文的供应链。他将展示如何从受信任的注册中心获取经过批准的包，在任何平台上进行固定和哈希验证，并通过不可绕过的 CI 门控来强制执行组织策略。
