内容来源：https://blog.google/innovation-and-ai/models-and-research/gemini-models/using-gemini-to-manage-farm/

内容总结：

密歇根奶农借助AI智能体实现高效管理：从数据孤岛到自动化运营

在利润微薄的奶牛养殖业中，数据分析是应对复杂生物与环境变量的关键。密歇根州奶农保罗·温德穆勒通过谷歌的Gemini 3.6 Flash模型，成功搭建了一套本地化多智能体系统，将原本耗时数小时的每日数据整合工作压缩至自动化流程，让自己能腾出时间专注于牧场经营。

保罗与妻子布里塔妮于2014年创办“梦想之风奶牛场”，从30头租赁奶牛起步，如今已发展为拥有260头荷斯坦奶牛的高度自动化牧场。作为2024年纽菲尔德国际农业学者，他将奶牛养殖视为技术挑战。然而，牧场产生的海量数据——包括牛群传感器项圈、气象站数据及牛奶质量记录——分散在不同软件系统中，导致他每天早晨需花费数小时下载文件、合并表格、计算绩效，严重挤占照料奶牛的时间。

为改变这一局面，保罗基于Gemini 3.6 Flash在Google Antigravity平台上构建了多智能体系统。该系统无需API或网络爬虫，仅通过本地文件夹接口工作。当CSV数据、收据照片、PDF发票等文件存入监控文件夹时，Gemini的多模态能力自动提取并整合视觉与数值信息，所有数据完全由保罗掌控。系统采用专业化智能体分工：调度智能体管理日常工作流，数据摄入智能体标准化原始文件，分析智能体评估生物与天气影响，报告智能体生成自然语言摘要。整个流程将原始文件自动转化为连贯的经营概览。

保罗的最终目标是让独立农户也能用上此类技术。此前，构建工具本身已是门槛，而大规模运行智能体系统的成本更是小型农场的难题。Gemini 3.6 Flash凭借其100万token上下文窗口、6.4万token最大输出及较前代约17%的token输出减少，显著降低了智能体循环的运营成本。该系统的核心指标是“静态可变利润”（SVM），通过锁定市场价格、只反映生物效率与运营健康度，排除了奶价与饲料价格波动的干扰，为决策提供可靠依据。

实际应用中，系统每日自动计算SVM并生成《牧场CEO简报》。例如，若每头牛SVM下降0.15美元，报告会精准溯源：干物质采食量减少（-0.08美元，因湿度降低采食量）、体细胞数上升（-0.04美元，暗示健康问题）、废弃奶增加（-0.03美元，两头牛进入治疗栏），并附上调整通风缓解热应激等可操作建议。由于系统基于本地文件运行，所有敏感数据均保留在牧场内。

保罗希望这一案例能启发更多小企业评估AI在日常工作流中的应用。通过多智能体架构与Gemini 3.6 Flash的低成本token效率结合，他打造了自动化运营层，彻底摆脱手动表格的束缚，真正实现“走出办公室，专注牧场发展”。

中文翻译：

密歇根州乳农借助Gemini Flash智能代理提升运营效率
乳品行业利润微薄，数据分析对管理复杂的生物与环境变量至关重要。现代农场产生的数据量庞大，如何快速分析并作出决策成为关键挑战。这正是保罗·温德穆勒选择Gemini 3.6 Flash的原因。
2014年，保罗与布列塔尼在密歇根州创办"梦想之风"乳业，初期仅租赁30头奶牛。12年间，该农场发展为高度自动化设施，饲养260头荷斯坦奶牛。作为2024年研究农业科技与人工智能的纽菲尔德国际农业学者，保罗将乳品养殖视为技术挑战。
其运营产生持续数据流：传感器项圈追踪每头奶牛，本地气象站记录气候指标，在线平台记录奶质与运输数据。
然而，这些系统形成了孤立的数据孤岛。每天清晨，保罗需花费数小时下载文件、合并电子表格并计算绩效，导致无暇照料奶牛。

将文书工作委托给AI代理
为解放时间，保罗利用Google Antigravity平台中的Gemini 3.6 Flash构建了本地多代理AI系统，整合孤立的农场数据并计算每日盈利。
该系统未使用API或网络爬虫，而是采用基于本地目录的文件接口。当CSV数据导出文件、纸质收据照片、PDF及发票存入监控文件夹时，Gemini的多模态能力将提取并合并视觉与数值指标，确保数据始终由保罗掌控。
专业化的多代理工作流程替代了冗长的提示指令，通过编排不同角色协同运作：

• 编排代理：管理整体日常流程
• 数据摄入代理：标准化原始文件（挤奶机器人导出数据、饲料记录）
• 分析代理：评估生物与气候影响
• 报告代理：生成清晰的自然语言摘要
  该系统自动将原始文件转化为连贯的业务概览。

为小企业创造公平竞争环境
保罗的终极目标是让独立农场主能像他一样，通过技术简化运营流程。对许多人而言，构建工具本身已是挑战，而后续的规模化应用更面临成本难题。
需要持续推理、解析和工具执行的日常智能代理工作流，对类似小型企业而言成本过高。Gemini 3.6 Flash使保罗的日常运营更具成本效益。
该模型专为高级推理、工具调用与编码设计，配备100万令牌上下文窗口与6.4万令牌最大输出。基准测试显示，相比Gemini 3.5 Flash，其输出令牌减少约17%，且单位输出令牌成本更低，显著降低了智能代理循环的运行成本。

以静态变量利润率衡量成功
系统优化目标直指保罗的核心指标：每日静态变量利润率。与传统"饲料成本收入比"等随奶价与饲料价格波动的指标不同，静态变量利润率将市场价格恒定化，从市场噪音中分离真实的生物与运营效率。这使智能代理系统基于可靠事实，为农场运营提供可执行的洞察。

智能代理工作流计算静态变量利润率

• 生物性能：数据摄入代理解析产奶量、成分（乳脂/蛋白质）、体细胞数与饲料数据
• 静态收入：基于历史支付标准（联邦第33号令一月价格）对实际乳固体产量应用固定价格
• 静态饲料成本：使用固定原料价格，仅反映实际饲料消耗量的变化
• 静态饲料成本收入比：静态收入减去静态饲料成本
• 可变成本：扣除每头奶牛支出（替换、繁殖、兽医用品），排除间接成本（人工、水电、折旧），确保静态变量利润率变化仅反映生物效率与健康状况

生成每日简报
计算数字仅是挑战的一半。忙碌的农场主无法每天凌晨3点解读电子表格。为节省时间，报告代理将结果转化为每日《农场CEO简报》，聚焦每日利润率驱动因素。例如，若静态变量利润率下降0.15美元/头，系统将精准定位原因：

• 干物质摄取量：-0.08美元（湿度降低饲料摄入）
• 体细胞数：-0.04美元（轻度升高预示潜在健康问题）
• 废弃牛奶：-0.03美元（两头奶牛进入治疗区）
  报告最终附上可执行建议，如调整通风缓解热应激。由于系统基于本地文件运行，所有敏感数据均保留在农场。

让独立企业超越电子表格实现规模化
保罗在Antigravity平台构建的智能代理系统，标志着小企业的重要转型。他希望通过这一案例启发并教导他人，如何评估在日常工作流中应用AI。通过将多代理架构与Gemini 3.6 Flash的低成本令牌效率相结合，他打造了自动化运营层，消除数小时的手动电子表格工作，使其能走出办公室，专注于业务增长。

英文来源：

How Gemini Flash agents are helping a Michigan dairy farmer
Dairy farming operates on tight margins, making data analysis essential for managing complex biological and environmental variables. Since modern farms generate massive volumes of information, the challenge lies in analyzing it quickly to make decisions. That’s why Paul Windemuller turned to Gemini 3.6 Flash.
In 2014, Paul and Brittany started Dream Winds Dairy in Michigan with 30 leased cows. Over 12 years, it grew into a highly automated facility milking 260 Holsteins. As a 2024 Nuffield International Farming Scholar studying ag tech and AI, Paul approaches dairy farming as a technological challenge.
His operations generate continuous data streams: sensor collars track each cow, a local weather station records climate metrics, and an online portal logs milk quality and shipments.
However, these systems operate in isolated software silos. Every morning, Paul spent hours downloading files, merging spreadsheets, and calculating performance. This kept him away from actually looking after the cows.
Delegating the office work to AI agents
To reclaim his time, Paul built a local, multi-agent AI system using Gemini 3.6 Flash inside Google Antigravity to integrate siloed farm data and calculate daily profitability.
Instead of APIs or web scraping, the system uses a local, directory-based file-interface. When CSV data exports or photos of papers receipts, PDFs, and invoices are saved to a monitored folder, Gemini’s multimodal power extracts and merges visual and numeric metrics — keeping data under Paul’s control.
A specialized multi-agent workflow replaces long prompts with specialized, orchestrated roles:

• Orchestrator: Manages the overall daily workflow.
• Ingestion Agents: Standardize raw files (milking robot exports, feed logs).
• Analysis Agent: Evaluates biological and weather impacts.
• Reporting Agent: Generates clear, natural language summaries.
  The system automatically transforms raw files into a cohesive business overview.
  Leveling the playing field for small businesses
  Ultimately, Paul’s ambition is to empower independent farmers with technology they can use to streamline their operations the same way he has been able to for his own business. Not only is building the tool a hurdle for many, but once they have them, operating at farm scale was a cost challenge that needed to be solved.
  The daily agentic workflows that require continuous reasoning, parsing, and tool execution have been too expensive for small businesses like his. Gemini 3.6 Flash helped make Paul's daily operation more cost-effective.
  Designed for advanced reasoning, tool use, and coding, Gemini 3.6 Flash features a 1 million token context window and a 64,000 token maximum output. Benchmark evaluations show it achieves roughly a 17 percent reduction in output tokens compared to Gemini 3.5 Flash, at a lower cost per output token, significantly lowering the cost of running agentic loops.
  Measuring success by Static Variable Margin
  The system is designed to optimize for Paul’s primary metric: Daily Static Variable Margin (SVM). Unlike traditional metrics like Income Over Feed Cost, which fluctuates with volatile milk and feed prices, SVM holds market prices constant, isolating the true biological and operational efficiency from market noise. This grounds the agentic system in a truth Paul can rely on to give him actionable insights for farm operations.
  How the agentic workflow calculates SVM
• Biological Performance: Ingestion agents parse milk yields, components (butterfat/protein), somatic cell count, and feed.
• Static Revenue: Applies fixed prices based on historical pay standards ( Federal Order 33 January prices) to actual milk solids produced.
• Static Feed Cost: Uses fixed ingredient prices to isolate changes in actual feed volume consumed.
• Static Income Over Feed Cost: Subtracts the static feed cost from static revenue.
• Variable Costs: Subtracts per-cow expenses (replacements, breeding, vet supplies) while excluding overhead (labor, utilities, depreciation), ensuring SVM changes reflect only biological efficiency and health.
  Generating the Daily Briefing
  Calculating numbers is only half the battle. Busy farmers don’t have time to decode spreadsheets every morning at 3:00 AM. To save time, the reporting agent translates results into a daily Farm CEO Briefing that isolates daily margin drivers. For example, if SVM drops by $0.15 per cow, it pinpoints specific causes:
• Dry Matter Intake: -$0.08 ( humidity reduced feed intake)
• Somatic Cell Count: -$0.04 (slight increase indicating potential health issues)
• Discarded Milk: -$0.03 (two cows entered the treatment pen)
  It ends with actionable recommendations, such as adjusting ventilation for heat stress. Since the system runs on local file exports, all sensitive data stays on the farm.
  Scaling independent businesses beyond the spreadsheet
  Paul’s agentic system built in Antigravity highlights a major shift for small businesses, which he hopes can inspire and teach others how they may evaluate using AI in their daily workflows. By pairing multi-agent architecture with the low cost token efficiency of Gemini 3.6 Flash, he built an automated operating layer that eliminates hours of manual spreadsheet work, allowing him to step out of the office, and focus on growing his business.