Agent正确率从26%飙到70%,最近爆火的EvoX有东西啊EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/BrhHRQRJe52nFJm7_wmzEw
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

前言

Agent 问世至今已有相当一段时间,大家想必对它们的运作机制已经相当熟悉了。

通常接到笔者布置的任务后,遇到简单的便独自处理;一旦发现任务较为繁杂,便会唤来几个子 Agent 充当帮手,协同完成工作。

常规做法是由总 Agent 统筹拆解并下发任务,子 Agent 各自负责检索资料、撰写文案或处理某个具体环节。待大家全都干完后,将成果汇报给总 Agent,由其复核一遍,最终汇总出一份完整的答复。

 

这种协作模式自然行得通。但问题在于,所有产出都需要总 Agent 重新解析一遍,底层的子 Agent 哪怕已得出精准结论,部分前提条件、数据及细节仍可能在汇总时出现遗漏。

举个例子,十个人各自解完十道题,却无法直接上交答卷,得先把答案口述给某位同学,再由这位同学重新誊写。就算题目全解对了,最终提交的答案,多少还是会有出现偏差的可能。

不过眼下,有个中国团队认为,这事儿未必非得设个“总管”来统一收取作业。

他们构想出一种名为“去中心化蜂群”的协作架构。即让多个 Agent 处于平等地位进行分工,每个 Agent 仅处理自身职责边界内的任务。任务完成后,程序会直接依据任务编号,将相应结果拼合起来,从而免去了总 Agent 重新理解和转述的环节。

Agent正确率从26%飙到70%,最近爆火的EvoX有东西啊

另外,他们还在探索让 Agent 将已验证可行的做法留存下来,并共享给其他 Agent 重复利用。后续若有了新的结果,这些经验教训还会被持续修改与更新。

这一发展路线被称为 AI for AI。通俗来讲,就是用 AI 来优化 AI。

近来,他们将上述能力整合成了一款普通用户也能上手的桌面 Agent:EvoX。


EvoX

首先让 EvoX 尝试开一间咖啡馆

这一次,我们暂且跳过理论概念。我直接向大家展示一下,我是如何用 EvoX 为某家咖啡馆策划开业方案的。

我事先整理了若干与咖啡店相关的资料,涵盖了店铺背景、菜单成本、人员排班、供应商报价以及场地限制,随后将这些信息一次性输入给 EvoX。

分配给它的任务同样不复杂:为该门店策划一套具备实操性的七天开业周方案。

EvoX 接收任务后,首先对整体事项进行了拆分。从界面展示可以看出,它调用了 5 个 Agent 来处理店铺的各个维度。

这五个 Agent 将同步开展工作,任意点击其中一个,都可以进一步查阅它被分配的具体任务以及当前的执行进度。

待它们分别完成各自的工作后,EvoX 最终为我生成了一份完整的开业周方案。

该方案包含了每日的工作事项、员工排班方式、菜单与活动的规划、宣传内容的发布时间、预算的具体去向,以及哪些环节尚存风险并需要店主亲自确认。

蜂群与 Sub-Agent 之间究竟存在哪些差异

看完上述流程,或许有人会疑惑:这不跟原来差不多吗?无非就是多找几个 Agent 一起来分担工作罢了?

事实上,两者的差异主要体现在最终的汇合方式上。

EvoX 团队为此专门设计了一组对照实验。他们收集了 563 道涵盖逻辑、数学和物理领域的题目,三种模式均采用同一模型 Claude Haiku 4.5,测试题目和评判标准也保持一致,唯一变量是 Agent 之间的协作方式。

结果显示,单 Agent 的准确率为 26.29%,Sub-Agent 模式达到 38.54%,而 EvoX 蜂群的表现最为突出,正确率攀升至 70.69%—70.87%。

更值得关注的是,Sub-Agent 在过程中其实已经答对了 373 道题,但经过总 Agent 汇总整合后,最终仅保留了 217 道,正确答案的留存率仅为 55.5%。

也就是说,很多题并不是子 Agent 不会做,而是在最后转交、理解和重写的时候,答案发生了变化 

回头再看前面的咖啡店 Case,关键就不仅仅是"同时调用了几个 Agent"。各自的职责划分、结果的存放位置,以及最终由程序直接完成汇合,这些才是蜂群与普通 Sub-Agent 模式真正产生差距的所在。

EvoMap 这个团队规模不到 20 人。

读到这儿,我不禁对 EvoX 背后的开发团队产生了兴趣,想知道他们究竟是怎样的一群人,以及为何会专注于这种协作模式。

答案多少有点意外。

EvoMap 创立于 2025 年的深圳,作为一家初创企业,其团队规模不足 20 人,且成员基本以 95 后和 00 后为主。在 AI 时代,人数少并不一定意味着处于下风。由于决策流程精简、试错成本低,加上利用 Agent 来倍增执行力,这种小型团队反而展现出更强的灵活性。

EvoX 正在做的事,本质上也是在回应团队内部遇到的困境:如何让多个 Agent 实现高效协同,以及如何把那些行之有效的流程固化下来。

在这个语境下,EvoMap 既代表了开发团队的名称,又指代他们构建的 Agent 经验网络;而 EvoX 则是面向普通用户开启这套能力的大门。

到了第二次,无需再次说明规则

在梳理完咖啡店的开业规划后,笔者能够将那些固定不变的原则交给 EvoX 进行存储。不仅如此,EvoX 在处理日常任务的过程中,也会自发地保留一些相关的记忆信息。

我尝试布置了一项新任务,要求它为同一家咖啡店策划一份为期三天的中秋活动方案。

此次,我仅提供了最新的预算、人员、库存与场地通知,并未重复阐述上一轮的经营规则。

任务一经启动,界面便迅速弹出提示,表明它回想起了上一轮的店铺数据,并调用了其中的四条记忆。

行文至此,EvoMap 的意图也就变得十分清晰易懂了。

这好比是一个专为 Agent 设立的经验交流论坛。当某个 Agent 成功完成任务后,会随手留下一篇攻略;而当其他 Agent 碰到相似困境时,便能先行查阅这份经验,随后结合自身当前的新状况加以运用。

正是由于沿用了先前的经验,本次的活动思路策划不仅耗时更短,细节也变得更加丰富。

把这套方法用进研究

EvoMap 团队更是将“蜂群+经验迭代”这一理念应用到了更为繁杂的研究任务中,打造出已开源的 AutoResearch。简而言之,AR 便是蜂群协同合作的成果。

在 AutoResearch 系统中,各个 Agent 分管构思方案、开展实验、执行测试与核查结果。行之有效的方案会晋级至下一轮,而未能成功的方案也不会被直接废弃,而是被当作后续改进的参考。

在官方披露的 Django 修复实验里,新增能力测试的通过率由 2/7 提升至 7/7,并且维持了 203/203 项回归测试的顺利通过。

人工智能优化人工智能:其应用场景将涵盖哪些领域

我们不妨预先设想一种未来或许会出现的场景。

在一个工作日的中午,商务区周边骤降暴雨。瞬间激增的订单导致店家备餐延迟,而外卖骑手不仅要涉水经过积水路段,还得在拥挤的取餐处和漫长的等电梯过程中耗费时间。

各个 Agent 能够各自探索不同的策略,例如压缩合并订单的区域范围、优先派送已准备好的餐品,或是绕开那些长期拥堵的取餐地点。

如果某种办法确实减少了骑手绕路,也让送达时间更加稳定,这套方法就可以连同当时的天气、订单密度和商家出餐情况一起保存。下一次,另一片区域遇到相似情况,Agent 找到这份经验,确认条件接近后再使用,不必从头试错 

相似的经验同样能应用于餐厅厨房的备餐调度、物流货运途中的临时改道,乃至更多普通大众日常所依赖的各类生活服务之中。

写在最后

如今,许多 Agent 依然在学习如何模仿人类去点击鼠标、填写表格以及浏览操作网页。这一阶段无疑具有价值,因为笔者现有的软件系统和工作流程本就是面向人类界面设计的。

然而,机器所擅长的领域与人类有所差异。它们能够同时并行处理多项任务,依据编号精准整合结果,并且能够留存可供机器直接读取与复用的经验。

未来的 Agent 将渐渐发展出专属的工具与行为模式,无需始终扮演一个替人类操作软件的数字打工者角色。

EvoMap 当前所探索的,正是这样一套专属于 Agent 自身的组织架构与经验沉淀平台。

人工智能正在实现自我优化。它不仅逐渐孕育出独有的合作机制,还着手将验证成功的策略进行存储、共享及持续迭代。

EvoX 与 AutoResearch 已使该领域展现出切实可见的成效,与此同时,EvoMap 正迈向下个十年的崭新起点不断前行。


▽▽▽

Source: https://mp.weixin.qq.com/s/BrhHRQRJe52nFJm7_wmzEw If there are copyright issues, please email us at 13951431913#139.com or submit a comment on the right.

Foreword

Agents have existed for quite some time now, and everyone is probably already very familiar with how they work.

Usually, after receiving a task assigned by me, if it is simple, it handles it on its own; once it discovers that the task is relatively complex, it brings in several sub-Agents as helpers to complete the work collaboratively.

The usual approach is for the main Agent to coordinate, decompose, and distribute tasks, while sub-Agents are responsible for searching materials, writing copy, or handling specific individual steps. After everyone finishes, the results are reported back to the main Agent, which reviews them once more and finally compiles a complete answer.

This collaborative model naturally works. But the problem is that all outputs need to be re-parsed by the main Agent, and even if the lower-level sub-Agents have already reached precise conclusions, some premises, data, and details may still be lost during aggregation.

For example, if ten people each solve ten problems but cannot directly submit their answers, they must first dictate the answers to one classmate, who then rewrites them. Even if all the problems are solved correctly, the final submitted answers may still have some chance of deviation.

However, at present, a Chinese team believes that this does not necessarily require setting up a "manager" to collect everyone's homework in a unified way.

They envisioned a collaborative architecture called a "decentralized swarm". In other words, multiple Agents collaborate as equals with divided responsibilities, and each Agent only handles tasks within its own scope of responsibility. After tasks are completed, the program directly pieces together the corresponding results according to task numbers, eliminating the need for the main Agent to re-understand and restate them.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

In addition, they are exploring ways for Agents to retain verified, workable practices and share them with other Agents for reuse. If new results emerge later, these experiences and lessons will continue to be modified and updated.

This development path is called AI for AI. Put simply, it means using AI to optimize AI.

Recently, they integrated the above capabilities into a desktop Agent that ordinary users can also use: EvoX.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

EvoX

First, have EvoX try opening a coffee shop.

This time, let's skip the theoretical concepts for now. I will directly show you how I use EvoX to plan an opening scheme for a certain café.

I had previously organized several materials related to coffee shops, covering the store's background, menu costs, staff scheduling, supplier quotes, and venue limitations, and then fed all of this information to EvoX at once.

The task assigned to it was also not complex: plan a practical seven-day opening-week scheme for that store.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

After EvoX receives the task, it first breaks down the overall work into parts. From the interface display, we can see that it invoked 5 Agents to handle different dimensions of the store.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

These five Agents work simultaneously; clicking any one of them allows you to view in detail the specific task assigned to it and its current execution progress.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

After they each complete their respective work, EvoX ultimately generates a complete opening-week plan for me.

The plan includes daily work tasks, staff scheduling methods, menu and event planning, publication times for promotional content, the specific allocation of the budget, and which parts still carry risks and need the shop owner to personally confirm.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

What exactly are the differences between a swarm and Sub-Agents?

After seeing the above process, some people might wonder: isn't this basically the same as before? It's just hiring a few more Agents to share the work, isn't it?

In fact, the differences between the two mainly lie in the final aggregation method.

The EvoX team designed a set of controlled experiments specifically for this. They collected 563 questions covering logic, mathematics, and physics. All three modes used the same model, Claude Haiku 4.5; the test questions and evaluation standards remained consistent as well, with the only variable being the way Agents collaborated with each other.

The results showed that the accuracy of a single Agent was 26.29%, the Sub-Agent mode reached 38.54%, and EvoX's swarm performed best, with the correct answer rate rising to 70.69%—70.87%.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

More notably, Sub-Agents had actually answered 373 questions correctly during the process, but after aggregation and consolidation by the main Agent, only 217 were ultimately retained, giving a correct-answer retention rate of only 55.5%.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

In other words, many questions were not ones that sub-Agents could not solve; rather, the answers changed during the final handoff, understanding, and rewritingEvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

Looking back at the previous café case, the key point is no longer just "how many Agents were invoked at the same time". The division of responsibilities, where results are stored, and the fact that the program directly completes the aggregation are where the swarm model truly differs from the ordinary Sub-Agent model.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

The EvoMap team has fewer than 20 members.

At this point in the article, I couldn't help becoming interested in the development team behind EvoX, wanting to know what kind of group they are and why they focus on this kind of collaborative model.

The answer was somewhat unexpected.

EvoMap was founded in Shenzhen in 2025. As a startup, its team has fewer than 20 members, and the members are basically mostly people born after 1995 and after 2000. In the AI era, having fewer people does not necessarily mean being at a disadvantage. Because the decision-making process is streamlined, trial-and-error costs are low, and Agents are used to amplify execution, this small team actually shows stronger flexibility.

What EvoX is doing is, in essence, also a response to difficulties encountered within the team: how to make multiple Agents collaborate efficiently, and how to solidify processes that have proven effective.

In this context, EvoMap both represents the name of the development team and refers to the Agent experience network they built; while EvoX is the gateway that opens these capabilities to ordinary users.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

For the second time, there is no need to explain the rules again.

After organizing the café's opening plan, I was able to give EvoX those unchanging principles to store. Moreover, in the process of handling daily tasks, EvoX also spontaneously retains some relevant memory information.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

I tried assigning a new task, asking it to plan a three-day Mid-Autumn Festival event scheme for the same café.

This time, I only provided the latest budget, staffing, inventory, and venue notices, without repeating the business rules from the previous round.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

As soon as the task started, a prompt quickly appeared in the interface, indicating that it recalled the store data from the previous round and invoked four of its memories.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

By this point in the article, EvoMap's intention becomes quite clear and easy to understand.

It is like an experience-sharing forum specifically set up for Agents. When an Agent successfully completes a task, it leaves behind a strategy note; when other Agents encounter similar dilemmas, they can first consult this experience, and then apply it in combination with their own current new circumstances.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

Precisely because it reused the previous experience, the event ideas for this round were not only planned in less time, but also with richer details.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

Applying this method to research.

The EvoMap team has further applied the concept of "swarm + experience iteration" to more complex research tasks, building the already open-sourced AutoResearch. Simply put, AR is the result of swarm collaborative work.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

In the AutoResearch system, different Agents divide responsibilities such as conceptualizing plans, conducting experiments, executing tests, and verifying results. Effective solutions are promoted to the next round, while unsuccessful solutions are not directly discarded but are used as references for subsequent improvements.

In the officially disclosed Django repair experiment, the pass rate for new-feature tests improved from 2/7 to 7/7, and all 203/203 regression tests continued to pass smoothly.

EvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

Artificial intelligence optimizing artificial intelligence: what application scenarios will it cover?

Let us first imagine a scenario that may possibly emerge in the future.

At noon on a workday, a sudden heavy rainstorm falls around a business district. The surge in orders causes delays in food preparation at stores, while delivery riders must not only ford through waterlogged streets but also spend time at crowded pickup points and in long elevator queues.

Agents can each explore different strategies, such as compressing and merging the area scope of orders, prioritizing delivery of already prepared meals, or bypassing pickup locations that are persistently congested.

If a method indeed reduces riders' detours and makes delivery times more stable, this method can be stored together with the weather, order density, and store output conditions at the time. The next time, when another area encounters a similar situation, an Agent can find this experience, confirm that the conditions are close enough, and then use it, without needing to start trial and error from scratchEvoX: Model accuracy leaps from 26% to 70%, gaining widespread attention

Similar experiences can also be applied to meal-preparation scheduling in restaurant kitchens, temporary route changes during logistics shipping, and many other everyday life services that ordinary people rely on.

In the end.

Nowadays, many Agents are still learning how to imitate humans by clicking the mouse, filling out forms, and browsing web pages. This stage undoubtedly has value, because my existing software systems and workflows were designed for human interfaces.

However, the domains in which machines excel differ from those of humans. They can process multiple tasks in parallel simultaneously, precisely integrate results based on numbers, and retain experiences that machines can directly read and reuse.

Future Agents will gradually develop their own dedicated tools and behavioral patterns, without always having to play the role of a digital worker operating software on behalf of humans.

What EvoMap is currently exploring is exactly an organizational architecture and experience-accumulation platform exclusively designed for Agents themselves.

Artificial intelligence is achieving self-optimization. It not only gradually develops its own collaborative mechanisms, but also begins to store, share, and continuously iterate on strategies that have been validated as successful.

EvoX and AutoResearch have already made the effectiveness of this field clearly visible, and at the same time, EvoMap is moving forward toward the starting point of the next decade.

▽▽▽

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry