
Research Efficiency Up 300%? DojoAgents Passes 1,000 GitHub Stars and Begins Rewriting the Financial Research Workflow
English Version First|中文完整版见后
Just over a week after its open-source release, DojoAgents surpassed 1,000 stars on GitHub.
Its significance is not that it can answer another stock-market question. The market already has plenty of financial chatbots capable of summarizing news, interpreting earnings reports, and drafting research notes.
What makes DojoAgents more interesting is that it attempts to rebuild the workflow behind financial research itself.
Instead of waiting for a user to ask one question at a time, the system is designed to continuously gather information, follow emerging clues, call analytical tools, test assumptions, and assist with portfolio and risk decisions.
This marks a broader shift in financial AI:
From generating answers to conducting research.
Over the past year, large language models have made news summaries, financial statement analysis, and report generation increasingly accessible. Once these capabilities become standard, however, a more practical question emerges:
What comes after the financial chatbot?
DojoAgents offers one possible answer.
AI should not merely produce a polished response. It should build an evolving understanding of markets and continue advancing research tasks within clearly defined boundaries of data access, tool permissions, risk controls, and human oversight.
From “Answering Questions” to “Conducting Independent Research”
Traditional financial AI works like a question-and-answer system.
You ask one question. It returns one answer.
Real investment research does not work that way.
Consider a seemingly simple question:
Nvidia fell sharply overnight. Will China’s optical-module, PCB, and semiconductor-equipment stocks be affected today?
A capable financial agent cannot answer this reliably from a single prompt.
It must first retrieve overnight market performance and relevant news, then determine whether the decline was caused by:
A change in AI demand
Earnings or guidance
Valuation pressure
Supply-chain developments
Policy risk
Broader market sentiment
Next, it must map the event across industries and markets:
U.S. AI chips → cloud infrastructure → servers → optical modules → PCB suppliers → semiconductor equipment
It then needs to identify which A-share and Hong Kong-listed companies may have direct or indirect exposure and determine whether the user’s existing portfolio contains related risks.
Finally, it must distinguish between:
Confirmed facts
Market expectations
Possible transmission paths
Unverified assumptions
This is not a task that one carefully written prompt can complete consistently.
It is a research chain composed of data retrieval, tool usage, continuous judgment, cross-market mapping, and repeated validation.
A chatbot focuses on:
How should I answer this question?
A financial agent must also determine:
What should I investigate next, and how should I verify it?
What Is DojoAgents?
DojoAgents is part of AlphaDojo, an open-source autonomous financial intelligence system developed by DeepWave Intelligence.
Its framework divides financial research into a set of tools that an agent can call independently, including:
Market scanning
News retrieval
Company analysis
Industry-chain mapping
Portfolio construction
Exposure diagnosis
Risk analysis
Performance attribution
These tools are then connected through an Agent Loop, creating a complete decision-support workflow.
The user no longer sees only a final conclusion.
The research process can also reveal:
Which data sources were called
Which clues were followed
How an event was mapped across industries
How a hypothesis was tested
Whether a simulated portfolio supported the conclusion
How the final judgment was formed step by step
This matters in finance because a conclusion without evidence is difficult to trust.
A research system becomes more useful when its process can be reviewed, challenged, and reproduced.
The Core Architecture: A Loop-Driven Cognitive Portfolio Agent
At the center of DojoAgents is what the project describes as a Loop-Driven Cognitive Portfolio Agent.
It is not merely a dashboard displaying market data. It is designed as an autonomous reasoning layer positioned alongside a user’s portfolio.
The engine covers four major dimensions.
1. Fundamental Data Intelligence
The system retrieves, structures, and cross-references raw information from multiple markets, including:
Price and volume data
Candlestick charts
Corporate financial statements
Company announcements
Market news
Industry information
The objective is not only to collect data, but to connect data from different sources and place it within a usable research context.
2. Higher-Order Analytical Reasoning
DojoAgents can examine factors such as:
Valuation ranges
Momentum indicators
Industry classification
Company fundamentals
Relative performance
Market structure
The agent attempts to move beyond surface-level price changes and identify the mechanisms that may be driving market behavior.
3. Cross-Market Strategy Mapping
Modern financial markets are increasingly interconnected.
A shift in U.S. technology stocks may affect Asian supply chains. A change in commodity prices may influence currencies, industrial companies, and downstream manufacturers. A policy signal in one country may change valuations in another market.
DojoAgents is designed to map these relationships across A-shares, U.S. equities, and Hong Kong stocks, helping users evaluate how one event may propagate through different assets and regions.
4. Dynamic Portfolio Management
The system can continuously analyze:
Risk exposure
Portfolio concentration
Net asset value movements
Sector allocation
Performance attribution
Potential correlation risks
Instead of reviewing a portfolio only when the user asks a question, the agent framework is designed to support ongoing diagnosis and analysis.
Through one conversational interface, users can move from global market observation to company research, cross-market mapping, and portfolio interpretation.
Real-World Use Cases: The Researcher’s Workflow Is Being Rebuilt
The most valuable DojoAgents scenarios are not necessarily the most spectacular.
They are the repetitive, time-consuming tasks that occupy much of a researcher’s working day.
Daily Market Briefing
A typical researcher begins the morning with a basic question:
What happened in global markets, and what deserves attention today?
Answering it manually may require checking multiple terminals, news services, industry classifications, and individual securities.
DojoAgents can enter a data-collection loop and call tools for:
Global indices
Industries and sectors
Individual stocks
Unusual price movements
Changes in trading activity
News catalysts
Geographic market performance
Information that might previously have required half an hour of manual collection can be reorganized into a structured market brief from a single request.
The value is not simply speed.
It allows the researcher to spend less time gathering information and more time evaluating why it matters.
News-Impact Analysis
When major news breaks, the first research question is often:
Which companies and assets may be affected?
A conventional chatbot may summarize the news. DojoAgents attempts to go further by decomposing the question and collecting information through market, company, news, and industry tools.
Take a hypothetical event such as “Meta sells excess AI computing capacity.”
The agent might examine:
Meta’s stock-price reaction
The company’s related announcements and news
Whether the action signals excess capacity, capital discipline, or changing demand
The possible transmission path across the value chain
That path might include:
AI computing → cloud services → servers → optical modules → data-center components
The researcher no longer needs to draw every transmission chain manually. The agent can propose the structure, retrieve supporting information, and identify areas that still require verification.
Portfolio Diagnosis
DojoAgents also supports multimodal interaction.
A user may upload a screenshot of portfolio holdings, after which the system can attempt to identify positions and analyze:
Sector concentration
Cross-market exposure
Factor overlap
Possible event sensitivity
Risk clustering
Potential diversification gaps
This does not replace regulated investment advice or human risk management. Its practical value lies in accelerating the first stage of diagnosis and helping researchers identify which exposures deserve closer examination.
Where Does the “300% Efficiency Improvement” Come From?
“Research efficiency increased by 300%” should not be interpreted as a universal result for every institution or research task.
Its logic comes from the way time is allocated in traditional financial research.
Researchers often spend a large portion of their working hours on operational tasks such as:
Collecting data
Cleaning fields
Converting formats
Searching news
Writing backtesting code
Adjusting parameters
Updating charts
Organizing reports
Tracking unresolved questions
A single research iteration can take days, leaving relatively little time for developing and challenging the underlying investment thesis.
The value of an agent system lies in compressing these repetitive stages.
A useful industry comparison comes from an AI-native investment-research system introduced by China Merchants Securities International. Its approach turned active stock selection into a more auditable and reviewable automated production line, compressing some qualitative research processes from weeks to hours.
DojoAgents is moving in a similar direction.
It turns fragmented information retrieval, analytical discussion, tool execution, and workflow tracking into a reusable human-agent research process.
In scenarios where a researcher previously spent three hours collecting and organizing information before completing one hour of substantive analysis, automating much of the preparatory work could produce a multiple-fold improvement in effective research output.
The real benefit is not that AI suddenly becomes three times smarter than the researcher.
It is that the researcher can dedicate a much larger proportion of time to the work that requires professional judgment.
Low Deployment Barriers: A Financial Agent in 10–15 Minutes
DojoAgents also lowers the technical barrier to testing an autonomous financial research workflow.
The project supports Python 3.11 and above and can be installed through the UV package manager. According to the project description, users can launch a local financial research agent in approximately 10 to 15 minutes, depending on the environment and model configuration.
It supports API-based models from multiple ecosystems, including:
OpenAI
Anthropic
Gemini
DeepSeek
GLM
It can also work with locally deployed models through frameworks such as Ollama and vLLM.
This flexibility matters for financial institutions.
Different teams may have different requirements for:
Model capability
Inference cost
Data privacy
Deployment location
Compliance
Latency
Auditability
A research framework that can switch between cloud APIs and local models is easier to adapt to different security and cost-control policies.
Compute Cost: The Final Barrier to Scaling Financial Agents
DojoAgents addresses the question of how financial research can be reorganized.
But a second obstacle appears when enterprises attempt to run agent workflows continuously:
Compute cost.
An Agent Loop does not make one model call and stop.
Every cycle may involve:
Scanning markets
Retrieving news
Reading corporate information
Mapping industries
Comparing assets
Diagnosing portfolios
Testing assumptions
Producing structured reports
Reviewing previous conclusions
Each step consumes Tokens.
When one analyst runs an occasional query, the cost may appear manageable. When a financial institution runs hundreds of agents across markets, portfolios, clients, and research teams, Token usage can expand rapidly.
The system may be capable of running.
The question is whether it is affordable to keep running.
Unexhub: Turning Token Capacity into a Market-Oriented Resource
This is where Unexhub enters the infrastructure layer.
Unexhub aims to transform Tokens from isolated resources sold by individual model providers into capacity that can be aggregated, distributed, routed, priced, and settled through a compute trading network.
Its proposed model includes three core mechanisms.
1. Token Aggregation and Distribution
Model and compute resources are distributed across large providers, regional computing centers, specialist AI companies, independent teams, and individual developers.
Smaller suppliers may possess capable models or idle inference capacity but lack:
Stable customer channels
Global distribution
Intelligent routing
Billing infrastructure
Market bargaining power
Unexhub brings these resources into a unified supply pool and distributes them to enterprise users and developers through standardized interfaces.
This gives suppliers an additional route to market while giving developers more model, endpoint, and pricing choices.
2. Faster Settlement
Traditional enterprise billing cycles can place substantial pressure on smaller suppliers.
Unexhub proposes settlement based more closely on actual API usage, enabling suppliers to view and receive revenue on a shorter cycle.
The objective is to reduce the period between providing model capacity and recovering funds.
Eligible early suppliers may also receive promotional commission rates as low as 0%, allowing them to test demand, build initial traffic, and validate their model’s commercial competitiveness with lower channel costs.
3. API Bidding and Dynamic Pricing
Instead of allowing one model provider to determine the only available price, Unexhub proposes a bidding mechanism in which effective API costs can respond to:
Available capacity
Real-time demand
Model performance
Latency
Service stability
Peak and off-peak periods
Competition among suppliers
Developers may obtain lower-cost capacity during periods of weaker demand, while suppliers may receive stronger returns when demand peaks.
Token pricing becomes a market result rather than only a number on an official price sheet.
From “Can It Run?” to “Can We Afford to Run It Continuously?”
DojoAgents and Unexhub operate at different layers of the same emerging ecosystem.
DojoAgents rebuilds the application layer.
It converts market scanning, news analysis, industry-chain mapping, portfolio diagnosis, and research reporting into an agent-driven workflow.
Unexhub addresses the infrastructure layer.
It attempts to reduce model-access costs through supply aggregation, intelligent routing, market pricing, and faster settlement.
One answers:
How should AI conduct financial research?
The other answers:
How can continuous AI research remain economically sustainable?
When a DojoAgents loop is operating in the background—repeatedly calling models for retrieval, interpretation, validation, and reporting—Unexhub’s compute trading network can route those requests toward more competitive model capacity.
Developers can pay according to actual Token consumption rather than purchasing unnecessarily large packages in advance.
Depending on model availability, supplier competition, route conditions, and promotional policies, selected services may be available at prices as low as approximately 30% of comparable market reference rates.
For an occasional query, the difference may appear modest.
For a financial institution running continuous multi-market research, portfolio monitoring, and risk-analysis agents, a lower unit cost can accumulate into substantial monthly and annual savings.
One Network, Two Transformations
Global financial markets have become a dense, interconnected network.
A semiconductor event in the United States can affect equipment companies in China, component manufacturers in Hong Kong, cloud-service valuations, currency expectations, and portfolio risk across multiple regions.
Research tools, however, are still frequently isolated.
Market data sits in one system. News sits in another. Portfolio information sits elsewhere. Backtesting tools and reports remain separated from the decision process.
DojoAgents is attempting to connect these research islands through a continuous agent loop.
Unexhub is attempting to make the compute required to maintain that connection more affordable.
Together, they represent two parts of the same transition:
Financial AI is moving from being able to run toward being affordable to run continuously.
The next generation of financial research systems will not be judged only by whether they can produce a persuasive answer.
They will also be judged by whether they can:
Show how the conclusion was reached
Continue updating that conclusion
Operate within data and permission boundaries
Support review and audit
Control long-term model costs
Scale from one analyst to an entire institution
DojoAgents may have passed 1,000 GitHub stars in just over a week, but the more important signal is not the star count.
It is the direction behind it:
The financial chatbot is giving way to the financial research agent—and the infrastructure required to keep that agent running is becoming part of the product itself.
研究效率提升 300%?DojoAgents 一周突破 1000 Star,金融智能体正在重写研究员工作流
中文完整版
开源仅一周多,DojoAgents 在 GitHub 上便突破了 1000 Star。
它受到关注,并不是因为市场上又多了一个可以回答股票问题的聊天机器人。总结新闻、解读财报、生成分析报告,这些能力已经逐渐成为金融大模型的标准配置。
DojoAgents 真正值得关注的地方,是它试图重新设计金融研究背后的整套工作流。
它不再只是等待用户提出一个问题,再生成一段看似完整的回答;而是能够持续采集信息、追踪线索、调用分析工具、验证假设,并辅助完成组合与风险研究。
这代表着金融 AI 正在发生一次重要转向:
从“生成答案”,走向“执行研究”。
过去一年,大模型已经把新闻摘要、财报解读和研报生成的门槛降到了很低。当这些功能逐渐成为标配,一个更加现实的问题也随之浮现:
金融聊天机器人之后,下一步是什么?
DojoAgents 给出了一种可能的答案。
AI 不应只生成一份语言流畅的结论,还应该在数据、工具权限、风险控制和人工监督的边界内,持续建立对市场的认知,并主动推进研究任务。
从“回答问题”到“独立推进研究”
传统金融 AI 的工作方式,更像一套问答系统。
用户问一句,它回答一句。
但真实的金融研究并不是这样运转的。
假设用户提出一个看似简单的问题:
英伟达隔夜大跌,今天 A 股的光模块、PCB 和半导体设备板块会不会受到影响?
一个合格的金融智能体,不能仅凭一个 Prompt 直接给出结论。
它首先需要获取隔夜市场走势和相关新闻,再判断英伟达下跌的原因究竟是什么:
AI 需求是否发生变化
财报或业绩指引是否低于预期
估值是否承受压力
供应链是否出现新变化
是否存在政策风险
还是单纯受到市场情绪影响
随后,它还需要把事件映射到不同市场和产业链中:
美国 AI 芯片 → 云计算基础设施 → 服务器 → 光模块 → PCB 供应商 → 半导体设备
接下来,它要找出 A 股和港股中哪些公司可能存在直接或间接敞口,并进一步判断用户的持仓中是否包含相应风险。
最后,它还必须区分:
已经确认的事实
市场的普遍预期
可能存在的传导路径
尚未得到验证的推测
这不是一个写得足够长、足够复杂的 Prompt 就能稳定完成的任务。
它是一条由数据检索、工具调用、持续判断、跨市场映射和反复验证组成的研究链。
聊天机器人更关心:
这个问题应该怎么回答?
金融智能体还必须进一步知道:
下一步应该查什么,又该如何验证?
DojoAgents 到底是什么?
DojoAgents 来自 AlphaDojo,这是由 DeepWave Intelligence 开发的一套开源自主金融智能系统。
它将金融研究拆分成一组可以由智能体独立调用的工具,包括:
市场扫描
新闻检索
公司分析
产业链映射
组合构建
持仓敞口诊断
风险分析
业绩归因
这些工具再通过 **Agent Loop(智能体循环)**连接起来,形成完整的决策支持流程。
用户看到的不再只是最终结论,还能够进一步了解:
智能体调用了哪些数据
它沿着哪些线索继续调查
事件是如何映射到不同产业的
某个假设经过了怎样的验证
模拟组合是否支持原有判断
最终结论是如何一步步形成的
这在金融场景中格外重要。
一份没有证据链的结论,很难获得真正的信任。
只有当研究过程可以复核、质疑和重复时,金融智能体才不只是一个内容生成器,而能逐渐成为可使用的研究工具。
核心架构:循环驱动的认知组合智能体
DojoAgents 的核心引擎被称为 Loop-Driven Cognitive Portfolio Agent,即循环驱动的认知组合智能体。
它并不是一个只负责展示行情数据的仪表盘,而是一层部署在投资组合旁边、能够持续推理和调用工具的智能体系统。
整个引擎覆盖四个主要维度。
一、基本面数据认知
系统可以获取、整理并交叉引用多个市场的原始信息,包括:
价格与成交量数据
K 线行情
公司财务报表
上市公司公告
市场新闻
行业信息
其目标不只是把数据堆积在一起,而是把不同来源的数据连接起来,放入可用于研究判断的语境之中。
二、高阶逻辑分析
DojoAgents 可以进一步分析:
估值区间
动量指标
行业分类
公司基本面
相对收益表现
市场结构变化
智能体尝试越过简单的价格涨跌,识别隐藏在市场波动背后的驱动机制。
三、跨市场策略演绎
今天的金融市场早已不是彼此独立的孤岛。
美国科技股的变化可能影响亚洲供应链;大宗商品价格可能同时影响汇率、工业企业和下游制造商;一个地区的政策信号,也可能重新改变另一个市场的估值逻辑。
DojoAgents 试图跨越 A 股、美股和港股,映射这些联动关系,帮助用户理解一个事件可能如何穿透不同资产和地区。
四、动态组合管理
系统还可以持续分析:
风险敞口
持仓集中度
净值变化
行业配置
业绩归因
潜在相关性风险
它不再只在用户提出问题时临时查看投资组合,而是试图支持更加持续的诊断和分析。
通过统一的对话入口,用户可以从全球市场观察,一路进入公司研究、产业链映射和组合风险判断。
真实场景:研究员的工作流正在被重新拆解
DojoAgents 最有价值的使用场景,未必是那些看起来最炫目的功能。
真正消耗研究员时间的,往往是每天反复执行、却又无法省略的基础工作。
每日市场概览
一个研究员每天早上打开终端,通常会先问:
昨晚全球市场发生了什么,今天有哪些事情值得关注?
如果完全依靠人工完成,研究员可能需要在多个终端、新闻平台、行业分类和个股页面之间来回切换。
DojoAgents 可以进入数据采集循环,调用相关工具整理:
全球主要指数
行业和板块表现
个股异常波动
成交变化
新闻催化因素
不同地区市场走势
过去可能需要半小时手动采集的信息,现在可以通过一句指令形成结构化的市场简报。
真正节省下来的,不只是时间。
而是让研究员把精力从“收集发生了什么”,转移到“判断为什么重要”。
新闻影响分析
重大新闻发生后,研究员的第一个问题通常是:
哪些公司和资产可能受到影响?
传统聊天机器人可能只会总结新闻,而 DojoAgents 会进一步拆解问题,通过市场、公司、新闻和行业工具收集信息。
以“Meta 出售闲置 AI 算力”这一假设性事件为例,智能体可能依次分析:
Meta 股价如何反应
公司公告和相关新闻表达了什么
这一行为反映的是算力过剩、资本纪律,还是需求变化
事件可能通过怎样的产业链路径继续传导
其分析路径可能包括:
AI 算力 → 云计算 → 服务器 → 光模块 → 数据中心零部件
研究员不必再从零手动画出每一条产业链传导图。智能体可以先提出结构、寻找支持信息,并标注仍需继续验证的环节。
持仓诊断
DojoAgents 还支持多模态交互。
用户上传一张持仓截图后,系统可以尝试识别相关标的,并进一步分析:
行业集中度
跨市场敞口
因子重叠
事件敏感性
风险聚集
潜在分散化缺口
它不能代替受监管的投资建议,也不能替代专业人员的最终风险判断。
它更现实的价值,是加快第一轮诊断,帮助研究员迅速找到最值得进一步检查的风险点。
“研究效率提升 300%”究竟是怎么算出来的?
“研究效率提升 300%”不应被理解为适用于所有机构和所有投研任务的绝对结论。
这一数字背后的逻辑,来自传统金融研究的时间分配方式。
研究员通常需要把大量工作时间花在操作性环节:
采集数据
清洗字段
转换格式
搜索新闻
编写回测代码
调整参数
更新图表
整理报告
追踪尚未解决的问题
一轮完整研究可能持续数天,真正留给策略逻辑、投资假设和反证分析的时间反而十分有限。
智能体系统的价值,就在于压缩这些重复步骤。
招商证券国际近期推出的 AI 原生投研 2.0 系统提供了一个值得参考的行业案例:它将主动选股流程改造成一条更可审计、可复核的自动化生产线,把部分原本需要数周完成的定性研究过程压缩到数小时。
DojoAgents 正在沿着相似的方向发展。
它把分散的信息收集、工具执行、分析讨论和流程追踪,重新整理为一套可复用、可迭代的人机协同研究框架。
例如,研究员过去可能需要用三个小时采集和整理材料,再花一个小时进行真正的研究判断。如果智能体能够自动完成大部分前置工作,同样时间内可投入实质分析的比例就可能出现数倍提升。
因此,所谓“300% 效率提升”,并不意味着 AI 的判断力突然变成了研究员的三倍。
它意味着研究员终于可以把更多时间投入那些真正需要专业经验与判断的工作。
10—15 分钟完成部署,降低金融智能体试用门槛
DojoAgents 的另一个特点,是技术部署门槛相对较低。
项目支持 Python 3.11 及以上版本,可以通过 UV 包管理器进行安装。按照项目说明,在环境和模型配置顺利的情况下,用户大约可以在 10—15 分钟内拉起一个本地金融研究智能体。
它支持多个 API 模型生态,包括:
OpenAI
Anthropic
Gemini
DeepSeek
GLM
同时也能通过 Ollama、vLLM 等框架接入本地部署模型。
这种灵活性对金融机构格外重要。
不同团队可能分别关注:
模型能力
推理成本
数据隐私
部署位置
合规要求
响应延迟
审计能力
一套可以在云端 API 与本地模型之间灵活切换的研究框架,更容易适配不同的安全和成本管理政策。
算力成本:金融智能体规模化的最后一道门槛
DojoAgents 回答了金融研究工作流应该如何重新组织。
但当企业准备让智能体持续运行时,另一个问题也会迅速出现:
算力成本。
Agent Loop 并不是调用一次模型就结束。
每一次循环都可能包含:
扫描市场
检索新闻
阅读公司信息
映射产业链
比较不同资产
诊断投资组合
验证研究假设
生成结构化报告
复核此前结论
每一个步骤都在持续消耗 Token。
当一名分析师偶尔进行一次查询时,成本可能并不明显。但当一家金融机构为不同市场、组合、客户和研究团队同时运行数百个智能体时,Token 用量会快速扩大。
系统已经“能跑”了。
接下来的问题是:
能不能长期跑得起?
Unexhub:让 Token 从厂商资源变成市场化资源
这正是 Unexhub 从基础设施层切入的位置。
Unexhub 希望将原本由不同模型厂商分别出售的 Token,转化为可以通过算力交易网络进行聚合、分发、路由、定价和结算的资源。
其模式可以拆分为三层核心机制。
第一层:Token 聚合与分发
模型和算力资源分散在大型平台、区域算力中心、垂直 AI 企业、独立团队和个人开发者手中。
中小型供应方可能拥有优秀模型或闲置推理容量,却缺少:
稳定客户渠道
全球分发能力
智能路由系统
完整计费基础设施
市场议价权
Unexhub 将这些资源聚合进统一供给池,再通过标准化接口分发给企业用户和开发者。
对供应方而言,这增加了一条商业化渠道;对开发者而言,则增加了更多模型、节点和价格选择。
第二层:更快结算
传统企业账期可能给中小供应方带来较大的资金压力。
Unexhub 希望按照真实 API 调用情况,更及时地记录和结算收益,缩短算力输出与资金回收之间的周期。
符合条件的早期供应方还可享受最低 0% 佣金的阶段性政策,在较低渠道成本下测试需求、积累流量,并验证自身模型的商业竞争力。
第三层:API 竞价与动态定价
Unexhub 不希望 Token 价格只能由单一厂商决定,而是提出 API 竞价机制,让实际价格可以根据以下因素动态变化:
可用算力容量
实时市场需求
模型性能
响应延迟
服务稳定性
高峰与低谷时段
供应方之间的竞争
开发者可以在需求较低时获得更优惠的算力,供应方也能够在需求高峰获得更合理的收益。
Token 价格由此不再只是一张官方价目表上的固定数字,而开始成为市场供需匹配后的结果。
从“能跑”走向“长期跑得起”
DojoAgents 与 Unexhub 处在同一套新生态的不同层级。
DojoAgents 重构的是应用层。
它把市场扫描、新闻分析、产业链映射、组合诊断和研究报告转化为智能体驱动的工作流。
Unexhub 重构的是基础设施层。
它希望通过供应聚合、智能路由、市场定价与快速结算,降低持续调用模型的成本。
一个回答:
AI 应该怎样做金融研究?
另一个回答:
持续运行 AI 研究系统,怎样才能在经济上可承受?
当 DojoAgents 的 Agent Loop 在后台持续检索、分析、验证并输出报告时,Unexhub 的算力交易网络可以尝试将这些请求路由至更具成本竞争力的模型容量。
开发者可以按照实际 Token 使用量付费,不必预先购买远超真实需求的套餐。
根据模型供应、节点竞争、路由条件和阶段性权益的不同,部分调用线路最低可享受约为市场参考价格 3 折的成本方案。
对于偶尔的一次查询,这种差异可能并不明显。
但对于持续运行跨市场研究、组合监控与风险分析智能体的金融机构而言,单位成本下降会逐步累积为明显的月度和年度预算节省。
一张网络,两场重构
全球金融市场早已变成一张高度密集的网络。
美国半导体行业的一次变化,可能同时影响中国设备厂商、港股零部件公司、云计算估值、汇率预期以及多个地区投资组合的风险。
然而,研究工具依然常常彼此割裂。
行情数据在一个系统,新闻在另一个系统,持仓信息又在其他平台;回测、报告与最终决策之间,也缺少连续的工作流。
DojoAgents 正在尝试通过持续运行的 Agent Loop,把这些研究孤岛连接起来。
Unexhub 则试图让维持这种连接所需的算力成本变得更加可承受。
两者共同代表了金融 AI 的一次重要转型:
从“系统能够运行”,走向“系统可以长期、规模化地运行”。
下一代金融研究系统的评价标准,将不再只是能否生成一份看起来专业的答案。
它还需要证明自己能够:
展示结论形成过程
持续更新原有判断
在数据和权限边界内运行
支持复核与审计
控制长期模型成本
从服务一名分析师扩展到服务整个机构
DojoAgents 开源一周多便突破 1000 Star,但真正重要的信号并不是 Star 数量本身。
而是它背后所代表的方向:
金融聊天机器人正在让位于金融研究智能体,而支撑智能体长期运行的算力基础设施,也正在成为产品能力的一部分。