新智元报道
3014.9亿斤。
国家统计局公布了夏粮数据，总产量3014.9亿斤，头一回站上3000亿斤，创历史新高。全年丰收已无悬念。
但在这个数字背后，有一场正在加速的技术变革——
一群科学家正在用AI，重新「设计」种子。
11月初，海南三亚，第八届国际农科院院长高层研讨会（GLAST）拉开帷幕。
全球四十多个国家的农科院院长、联合国粮农组织（FAO）和国际农业研究磋商组织（CGIAR）的代表齐聚一堂，核心议题只有一个：科技创新如何塑造未来农业。
在大会边会上，中国农业科学院作物科学研究所的付俊杰研究员做了一个10分钟的演讲，展示了一个正在改变育种底层工作流的系统——
农科发现·探索者，国内首个农业科研智能体。技术合作方是华为云。
要理解这套系统解决的问题，先看清传统育种科研需要突破的点在哪里。
我国育种在上下游联动不充分。上游的基因发掘、多组学分析、基因编辑已经达到分子级精度，但下游从多基因聚合到品种应用的整条研发链路，仍然很大程度依赖反复试验与多轮筛选，周期长达10年以上。
矛盾集中在三个地方：
成功率低。农业生物性状受多基因网络调控，关键基因发现长期依赖研究者的经验与多轮验证。
数据利用难。核心物种的基因组信息、百亿条生物学数据散落在不同数据库中，格式不统一，工具不互通。海量科研数据长期处于「不联动」状态，挖掘利用率低。
流程碎片化。从文献检索到数据分析，再到育种方案设计，每个环节彼此联通不够。一轮完整的育种方案研发，可能需要3到6个月起步。
2025年，国务院发布《关于深入实施「人工智能+」行动的意见》，明确提出「加快农业数智化转型升级」。
针对上述瓶颈，华为云与中国农科院给出的技术路径不是做一个「农业版ChatGPT」，而是搭建一套能自主规划、执行和验证科研任务的智能体系统。
从底层往上看，农科发现·探索者的技术栈分为四层：算力底座、模型中枢引擎、智能体协同与开放生态，以及承载这一切运行的智能体平台。
第一层：算力底座。基于华为云超智融合基础算力资源，通过ModelArts平台实现各类模型的预训练、后训练以及训练后的推理部署；通过MaaS平台调度三方开源LLM模型，科研模型则统一纳管在AI资产广场，共同支撑智能体系统中多种模型的业务应用。
对科研机构来说，实际的价值有两点：
一是昇腾算力对大模型预训练、后训练和推理的支撑已经成熟，科研模型不必再为算力适配单独投入；
二是云的统一调度能力——一套多智能体系统里同时运行着通用大模型、科研专用模型和上百种分析工具。在同一个体系内统一调度，整套系统才能稳定运转。
第二层：模型中枢引擎。系统核心是基于华为盘古基础大模型，经过海量农科文献知识预训练和后训练的农业领域基础大模型系统，模型执行自然语言理解与交互、多步推理与科研任务规划等不同任务。
一类是「理解」——科学家用自然语言提出问题，系统理解意图、检索文献、生成知识；另一类是「推理」——系统规划实验路径、做基因功能推断、预测性状表达。不同任务协同调度，在任务准确性和幻觉控制上都更可控。
第三层：多智能体协同与开放生态。这是科学家直接使用的层面。
系统采用MCP（Model Context Protocol）架构，由模型中枢引擎负责总体调度。
科学家提出研究问题，比如「找到控制水稻产量的关键基因组合」，中枢引擎将问题拆解为子任务，分派给不同专业智能体：「知识智能体」从知识文库中定位相关研究成果；「数据智能体」接管组学数据分析；「分析智能体」完成统计建模和关联分析；「模型智能体」执行性状预测和育种方案设计。
当中枢引擎接收复杂指令时，一句话触发的是超过千步的计算流程：自主拆解任务，调度知识、数据、分析等智能体协同执行，完成从文献证据汇聚、跨组学关联到网络构建与推断的智能化整合。
这一层还有一个决定系统能走多远的设计：智能体的技能（Skill）不是由工程师写死的，而是由农科院的科学家自己写。
这是真正意义上的农业领域共创：华为云提供框架和运行环境，领域知识由最懂这个领域的人来写。
这些Skill、工具、MCP、提示词和模型，统一沉淀到华为云AI资产广场进行纳管和分发，并与智果AgentArts等开发和运行平台打通。
一个研究所写出的水稻性状分析Skill，另一个研究所可以直接调用，不必从头再来。
再往外看一层：农科发现·探索者所在的科学计算专区，是华为云行业AI梦工厂七大专区之一。
截至目前，行业AI梦工厂已汇聚200多家核心伙伴、1500多个行业资产和1000多个落地项目，覆盖智慧医疗、具身智能、科学计算、智慧金融、智能制造、AI硬件和政务七个领域。
第四层：智果AgentArts智能体平台。上面这些智能体、Skill和资产要稳定、安全地运转，需要一个运行与治理底座。
智果（AgentArts）是华为云的企业级智能体运行平台，在这套系统中承担多智能体架构中大模型和工具的运行、调度中枢。
它提供四项关键能力：
一是安全隔离的运行环境。通过Runtime、Sandbox和身份认证机制，科研院所和企业可在同一套系统上多组织协同研究，这是前提条件。
二是长短期分层记忆。科学家每完成一次研究，GWAS分析结果、候选基因注释、QTL定位信息、田间验证反馈都会结构化沉淀，后续研究可按需检索，避免重复分析。
三是统一的工具接入。通过MCP Gateway将100+种生物信息学分析工具、基因组数据库封装为标准化可调用能力，支撑跨学科智能体协同。
四是全链路可观测。每个智能体的调用链、任务执行耗时、异常信息全程可追踪，科研分析过程可复核、可解释。
简言之，就是给每个科学家配了一支7×24小时不间断工作的AI科学研发团队。
以上是这套系统的通用科研范式。
农业涉及多个领域，包括作物、畜禽、水产、微生物、林草等——同一套智能体架构落到不同物种育种上，需要的数据、领域工具和具体验证方式不同。
范式要落地，靠的是协作。
中国农业科学院一侧，生物技术研究所牵头，联合作物科学研究所和北京畜牧兽医研究所构成了「发现→验证→应用」的完整链路：
生物技术研究所负责AI算法引擎和基因编辑等底层技术研发，是科研能力的源头；作物科学研究所提供主粮应用与应用基础研究，确保AI输出的方案能在真实条件下兑现；北京畜牧兽医研究所负责向动物育种方向拓展，完成产业端的应用验证。
三所的分工逻辑是：懂技术、会育种、能落地，形成不可拆分的闭环。
华为云一侧，提供全栈技术底座：昇腾算力 + ModelArts + MaaS（三方开源模型）+ 盘古大模型 + 智果AgentArts智能体平台 + AI DataLake。
从底层算力，到AI DataLake的数据治理，到AI资产广场的生态开放，再到智果AgentArts承载的科研智能体场景落地，这是一条端到端的链路，每一环都在同一个平台体系内。
数据治理是其中容易被忽略的一环。从表型和多组学数据从入库、清洗到训练、推理，全程在同一平台内完成，按项目和授权范围隔离。多个研究所、多家企业能在同一套系统上协同研究，靠的正是这个基础。
双方在2025年6月的华为开发者大会（HDC）上全球首发了这套系统，2026年6月正式揭牌「AI农业联合创新实验室」。
系统上线后，一个被反复提及的机制是「数据飞轮」。
这个概念需要拆开来看。
传统科研的数据积累是线性的：做一次实验，产出一批数据，发一篇论文，很多数据沉淀在硬盘里利用效率不高。
农科发现·探索者的做法不同——
每一次研究的原始数据、分析结论、方法路径都会回流到系统中，经过自动清洗、质量检测和结构化入库，变成可复用的数字资产。系统通过RLAIF（基于AI反馈的强化学习）机制持续优化模型表现。
截至目前，系统已积累3000多次真实农业研发全流程实验数据。
大模型最不缺的是文献和公开数据库，最缺的是湿实验数据：真实实验室和田间产出的一手结果。
农科发现·探索者积累的这3000多次实验数据，是在同一套系统里、按统一规范产生，并经过清洗和标准化的全流程数据，从问题提出、方案设计、数据分析到田间反馈，链条完整。这是比较难获取的数据。
这意味着系统不是一个静态工具，每一次使用都在优化下一次的输出质量。
飞轮转起来之后，AI育种的说服力最终取决于一个问题：设计出的方案，能不能真正改良目标性状，并在大田有很好的表现？
目前公开披露的核心案例：研究团队的目标是培育抗倒伏、产量不减的水稻新材料。传统路径下，仅筛选候选基因就可能就需要数月。
使用农科发现·探索者后，智能体自主完成文献检索、基因数据挖掘、最优基因组合锁定，直接生成育种方案。育种方案研发周期从3到6个月压缩到数周；24小时内完成超千步骤的数据分析，节省90%分析时间；人力效率提升10倍以上。
方案没有停留在理论层面。团队拿着AI设计的方案回到田间进行多个实验靶点的并行验证，分析结果：新材料抗倒伏能力显著提升，产量不受影响。
付俊杰研究员在GLAST的演讲里有一句话说得特别好：「所有的计算，最终都回到了田间。」
技术验证之外，产业端的进展尤为值得关注，农科发现·探索者正在赋能涉农领军企业，在表型分析、育种基因挖掘、遗传评估、生产管控等产业链条环节提高效率和降低成本。
今年6月，中国农科院生物技术研究所与华为揭牌「AI农业联合创新实验室」，农科发现·探索者同月首批入驻华为云行业AI梦工厂科学计算专区。北京丰台区挂牌了「AI育种大模型」创新中心，目标是把这套能力做成种业的数字底座。
7月世界人工智能大会（WAIC）上，农科发现·探索者获评「高价值应用场景典型案例」。
而现在，它到了GLAST——全球四十多个国家的农科院院长同桌而坐的场合。
对许多全球南方国家的农业科研院所而言，也同样需要这样一套AI赋能的科研系统，基于开放的智能体生态提供了一种实现路径：算力、模型和运行平台已经就绪，各国科学家只需把自己作物、自己地区的育种知识写成Skill，接入同一套系统。
中国在智能育种方向上积累的技术方案与产业落地经验，头一回在国际同行面前亮相。
站在三亚的椰风里回看这一切，有一种奇妙的时空感。
全世界都在讨论AI如何写诗、画画、剪视频，但有一群科学家和工程师，选了一条更安静的路——
在极端气候频发的当下，用AI加速育种研究，让种子更好地适应环境变化，让大家吃得更营养、更健康。
2024年诺贝尔化学奖给了预测蛋白质结构的AlphaFold，让全世界看到AI在生命科学领域的爆发力。
而在中国大地上，AI正在把这种能力从实验室推向田间地头。
这，或许才是AI最该去的地方之一。
编辑：所罗门
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
30.149 billion pounds.
The National Statistics Office published data on summer grains, which produced a total of 30,149 million pounds, making the first stop a new 300 billion pounds. The year-round harvest was over.
But behind this number is a technological change that is accelerating.
A group of scientists are using AI to "design" seeds.
At the beginning of November, the eighth International Academy of Agriculture (GLAST) high-level seminar opened in Hainan Sanya.
The presidents of agricultural institutes in more than 40 countries around the world, representatives of the Food and Agriculture Organization of the United Nations (FAO) and the Consultative Group on International Agricultural Research (CGIAR), have gathered together on one of the core issues: how science, technology and innovation shape future agriculture.
On the margins of the General Assembly, a 10-minute speech by a fellow from the Crop Sciences Research Institute of the Chinese Academy of Agricultural Sciences, Fu Junjie, showcased a system that was changing the bottom workflow in breeding --
AI agent. The technical cooperation component is Chinese.
To understand the problems addressed by the system, we need to see where traditional breeding research needs to break.
We have inadequate breeding links up and down. Upstream genetic excavations, multi-group analysis and genetic editing have reached molecular-grade accuracy, but the entire R & D chain downstream, from multi-genetic aggregation to variety applications, continues to rely heavily on repeated trials and multiple rounds of screening, with cycles of more than 10 years.
The contradictions are concentrated in three places:
success rate low. Agricultural biomorphology is regulated by a multigenetic network, and key genetic discoveries have long relied on the experience and multiple rounds of validation of researchers.
Data use is difficult. Genomic information on the core species, the tens of billions of biological data scattered in different databases, the formats are not uniform and the tools are not interoperable. Mass scientific data have been "unconnected" for a long time, with low utilization of excavations.
Process fragmentation. From literature to data analysis, to breeding programme design, each link is not sufficiently connected. A full cycle of seed development may take three to six months.
In 2025, the State Council, publish, issued an opinion on the further implementation of the "Advanced Intelligence Plus" initiative, in which it explicitly stated that "accelerating the upgrading of the intellectual transformation of agricultural numbers".
In response to the above-mentioned bottlenecks, the technical path given by Hua to Yun and the Chinese Agricultural Institute is not an "agricultural version of ChatGPT", but rather a system of AI agent that can plan, implement and validate research tasks on its own.
From the bottom up, the agronomy found that the explorer ' s technology stacks were divided into four layers: the algorithmic base, the model hub engine, AI agent synergetic and open ecology, and the AI agent platform, which carried all of this.
First floor: the bottom of the force. Based on China ' s basic computing resources for cloud super-intellectual integration, pre-training, post-training and post-training reasoning deployment of models are achieved through the ModelArts platform; the tripartite open-source LLM model is deployed through the Maas platform, and the scientific model is harmonized at the AI Asset Square to support the operational application of multiple models in the AI agent system.
For scientific institutions, there are two practical values:
The first is that calculus support for LLM pre-training, post-training and reasoning has matured and that scientific models need no longer be equipped with individual inputs for calculus;
The second is the unified movement capability of the clouds, which operates simultaneously in a multi-AI agent system with generic LLM, specialized scientific models and hundreds of analytical tools. Unified movement within the same system allows the entire system to function steadily.
Second level: Model central engine. The core of the system is based on LLM, which is based on China ' s ancient foundation, the LLM system for the basic agricultural field, which has been pre- and post-trained in the knowledge of mass agricultural literature, and modeling different tasks such as natural language understanding and interaction, multistep reasoning and scientific mission planning.
One is "understanding" — scientists ask questions in natural languages, systematically understand intent, retrieve literature, generate knowledge; the other is "adjection" — systematic planning of experimental paths, genetic function extrapolations, predictive expression. The coordinated movement of different tasks is more manageable in terms of mission accuracy and hallucination control.
Third floor: Multi AI agent synergetic and open ecology. This is the level of direct use by scientists.
The system follows the MCP (Model Context Protocol) structure, with the model hub engine responsible for the overall movement.
Scientists raised research issues, such as "Finding critical gene combinations to control rice production", the central engine decomposes the problem to subtask and assigns it to different professions, AI agent "Knowledge Z42Z" to locate relevant research results from the knowledge base; "Data AI agent" takes over the cluster data analysis; "Analyzing AI agent" completes statistical modelling and correlation analysis; and "Model AI agent" implements profiling and breeding program design.
When the central engine receives complex instructions, the first sentence triggers more than a thousand steps of the computational process: self-dismantling, co-execution of AI agent in the movement of knowledge, data, analysis, etc., completes the intellectual integration from the collection of documentary evidence, cross-grouping to network construction and extrapolation.
There is also a design that determines how far the system can go: the skills of AI agent (Skill) are written not by engineers, but by scientists at the Agricultural College.
This is the true agricultural sector: China provides the framework and operating environment for the cloud, and knowledge of the field is written by those who know it best.
These Skills, Tools, MCP, prompt and Models are integrated into China to provide a navigation and distribution to CloudAi Asset Square and connect with platforms developed and run by Jigo Agent Arts and others.
One research institute wrote a rice-type analysis of Skill, while another research institute could be called directly, without having to start over.
On the outside, the Agricultural Research Explorer's Scientific Calculator is located in one of China's seven Ai Dreams.
To date, the Industry AI Dream Plant has brought together more than 200 core partners, 1,500 industry assets and 1,000 landing projects covering seven areas: smart medicine, smartness, scientific computing, smart finance, smart manufacturing, AI hardware and government.
Level 4: Agent ArtsZ42Z platform. These AI agent, Skill and assets above need a base of operation and governance to operate in a stable and safe manner.
AgentArts, an enterprise-level AI agent operating platform for Chinese clouds, is responsible for the operation and dispatch of LLM and tools in the multi-AI agent architecture.
It provides four key capabilities:
One is a safe and isolated operating environment. Through Runtime, Sandbox and the Identification Mechanism, research institutes and enterprises can organize multiple research on the same system, which is a prerequisite.
The second is short- and long-term stratification memory. For each completed study, the results of GWAS analysis, candidate gene notes, QTL positioning information, field validation feedback are structured to settle, and follow-up studies can be retrieved as needed to avoid duplication of analysis.
The third is uniform access to tools. Supporting interdisciplinary AI agent synergies through MCP Gateway, the 100+ bioinformatics analysis tool and the encapsulation of the genomic database as a standardized callable capability.
Fourth is the full chain. The chain of call for each AI agent, the time taken to carry out the mission, the entire process of unusual information can be tracked and the scientific analysis process can be reviewed and explained.
In short, each scientist was assigned an AI scientific research and development team that worked 24 hours a day.
These are the general scientific paradigms of the system.
Agriculture covers a wide range of areas, including crops, livestock, aquatic, micro-organisms, forest grasses, etc. — the same AI agent structure falls on different species and requires different data, field tools and specific validation methods.
The paradigm depends on collaboration.
On the side of the Chinese Academy of Agricultural Sciences, the Institute of Biotechnology, led by the United Crop Science Institute and the Beijing Livestock and Veterinary Institute, constitute a complete link to the "detection and validation applications":
The Institute of Biotechnology is responsible for the research and development of lower-level technologies, such as the AI algorithm engine and genetic editing, and is the source of scientific capabilities; the Crop Science Institute provides basic research on staple food applications and applications to ensure that the AI-output programme is delivered under real conditions; and the Beijing Livestock and Veterinary Institute is responsible for the expansion of animal breeding directions and the validation of industrial applications.
The logic of the division of labour is that it is technical, breedable, landable, and forms a closed ring that cannot be split.
China is on the side of the cloud and provides the base of all-canton technology: + ModelArts + MaaS (triangular open-source model) + LLM + Agent Arts Z42Z platform + AI DataLake.
From bottom-up algorithms, to the data governance of AI DataLake, to the ecological opening of AI Asset Square, to the site of the scientific AI agent site carried by Jiguo AgentArts, an end-to-end link, each in the same platform system.
Data governance is an easily neglected link. From tabular and multi-group data entry, cleansing to training, reasoning, the entire process is completed within the same platform and isolated by project and mandate. This is the basis on which a number of research institutes and enterprises can work together on the same system.
The system was launched globally at the China-Developer Conference (HDC) in June 2025 and the "AI Joint Agricultural Innovation Laboratory" was officially unveiled in June 2026.
Once the system is online, one of the mechanisms repeatedly mentioned is the Data Flying Wheel.
This concept needs to be looked at.
The accumulation of data from traditional research is linear: an experiment, the production of data, a paper, and the inefficient use of data in hard drives.
There's a difference in the approach of the agro-science discoverer.
The raw data, analytical findings and methodological pathways of each study flow back into the system and are automatically cleansed, quality tested and structured into a reusable digital asset. The system continuously optimizes model performance through the RLAIF mechanism.
To date, the system has accumulated more than 3,000 real agricultural research and development process-wide experimental data.
LLM is the least lacking in literature and open databases, and the least lacking is data on wet experiments: real laboratory and field outputs.
The more than 3,000 experimental data accumulated by the Agricultural Section found by explorers were produced in the same system, by uniform norms, and were cleansed and standardized with full process data, ranging from problem formulation, programme design, data analysis to field feedback, and chain integrity. This is more difficult to obtain.
This means that the system is not a static tool and that each use optimizes the quality of the next output.
When the wheel turns around, the persuasiveness of AI breeding ultimately depends on the question: can the design of the formula really improve the target character and perform well in Daejeon?
Core cases of public disclosure: The research team ' s objective is to develop new, inverted and productive rice materials. Under the traditional path, the screening of candidate genes alone may take months.
Using the Agricultural Research Explorer, AI agent autonomously completes the literature search, genetic data mining, optimal combination locking, and directly generates breeding programmes. The research and development cycle of the breeding programme has been reduced from three to six months to several weeks; data analysis of over 1,000 steps has been completed within 24 hours, saving 90 per cent of analytical time; and human efficiency has increased more than tenfold.
The programme does not remain theoretical. The team returned to the fields with the AI-designed programme for parallel validation of multiple experimental target points and analysed the results of a significant increase in the invertability of new materials, with no impact on production.
In his speech at the GLAST, Fellow Fou Joonjie said, "All calculations eventually go back to the fields."
In addition to technical certification, progress at the industry level is of particular interest, with agro-studies finding that explorers are empowering agro-collar enterprises to improve efficiency and reduce costs in industrial chains such as surface analysis, genetic extraction, genetic evaluation and production control.
In June this year, the Institute of Biotechnology of the Chinese Institute of Agricultural Sciences and China unveiled the "AI Joint Agricultural Innovation Laboratory" and the first explorers of the same month were admitted to the Ai Dream Factory Science Calculator Area in China. The "AI breed Z43Z" innovation centre in the Tondai region of Beijing was established with the aim of making this capability the digital base of the seed industry.
At the World Congress of Artificial Intelligence (WAIC) in July, agroecologists discovered that explorers had been awarded "exact examples of high-value applications".
And now it is sitting at the table of the presidents of agricultural schools in over 40 countries around the world.
Such an AI-enabled scientific system is equally needed for agricultural research and research institutes in many countries of the global South. An open AI agent ecology provides a path to realization: computing, modelling and operating platforms are in place, and scientists in countries need only to write their own crops, their own areas, and their breeding knowledge, into the same system.
For the first time, China's technological programmes and industry's experience in the direction of intelligent breeding have appeared before its international counterparts.
It's a wonderful sense of time and space to look back in Sanya's coconut wind.
The whole world is talking about how AI writes poems, paintings, clippings, but a group of scientists and engineers have chosen a quieter path.
In the immediate aftermath of extreme weather events, the use of AI to accelerate breeding research has made seeds more adaptable to environmental change and more nutritious and healthy.
The 2024 Nobel Prize for Chemistry awarded AlphaFold, which predicts protein structures, to show the world the explosive power of AI in life sciences.
On Chinese soil, AI is pushing this capability from laboratory to field.
This, maybe, is one of the best places for AI.
Editor: Solomon
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.