跪拜 Guibai
← Back to the summary

EvoX Beta Ships a Swarm Mode That Hits 71% Accuracy by Splitting Agent Contexts

Hello everyone, I'm Brother Er.

For me, the two best Agent products at hand are Codex and Claude Code.

But for many of you, the barrier to entry for these two is a bit high, so every day in the group chat, people are discussing trying this Agent today and testing that Agent tomorrow.

I've also tried many other Agent products, and today I want to give a special recommendation to EvoX.

The interface is very clean, a geek style that I like.

To be honest, EvoX is indeed different from the Agent products I've used before. A group of AIs break down tasks themselves, work separately, and then aggregate the results. The EvoMap team calls this the 'Swarm Mode'.

It's not about launching more Agents, but about breaking tasks down more finely to reduce the complexity of a single task; at the same time, each Agent only focuses on its own thing, avoiding context pollution.

There is also a 'Networked Self-Evolution' mechanism, where each Agent can inherit the experience accumulated by other Agents on the EvoMap network. It's like having a top student cloned for you by default.

And you can use it just by registering, no invitation code needed. The built-in models include DeepSeek V4 and Kimi K3. Well, the Beta version just launched and is free to use directly. I can only say that product people nowadays are really thoughtful, with a full sense of service 😄.

Of course, if you have an API Key, you can also configure it directly.

I configured DeepSeek V4 Flash, GLM-5.2, and Qwen3.8 Max for it respectively. Under the same Harness architecture and task conditions, the final deliverables far exceeded my expectations.

And what's more, if you're smart, you can also directly use their built-in top-tier overseas models, such as Opus 4.8 and GPT-5.6 Sol.

Official website download: https://evomap.ai/evox/beta

Available for both Mac and Windows, the beta version just launched on August 9th, a brand new product.

01. First Experience with EvoX

EvoX divides its capabilities into three directions.

Chat is for Q&A and research, favoring structured output, with answers automatically organized into tables and charts.

Cowork is for collaborative delivery in office scenarios, such as summarizing meeting minutes, organizing documents, connecting domestic and international IMs to automatically aggregate information, and automatically executing long-running task flows, freeing people from repetitive labor.

Code is for development scenarios: reading projects, modifying code, running tests.

For Agent initialization, what I care about most is the model configuration.

If you have registered for EvoMap before, you can also select the built-in models in the backend, supporting DeepSeek V4 Flash, Opus 4.8, GLM-5.2, GPT-5.6 Sol, etc.

Every user receives trial credits just by registering.

Of course, it also supports configuring our own DeepSeek V4 Flash, GLM-5.2, Qwen3.8 Max, and local models.

The configuration process is simple.

Go to Settings, select AI Services.

Add a service, select Other Services, choose DeepSeek.

Fill in the API key, click [Test and Save] and you're done.

02. Chat Test

After configuring DeepSeek V4 Flash, let's first try out the Chat mode.

Create a new session and enter the prompt:

Help me design an AI style preset. Theme: [Programmer who likes to write AI-level review content]. Please write two sets of instructions around the theme: 'Naming Style' and 'Tone Style', and confirm saving with a style preset card.

Choose the DeepSeek V4 Flash model we just configured.

Expand the card to see the naming style and tone style.

Then go to Settings, select Personalization, and click [New Style].

Copy and paste the naming style and tone style here.

EvoX's future behavior will then better align with our personal needs.

The response speed is also very smooth.

DeepSeek V4 Flash is fast itself, and the response experience within EvoX is very good; basically, content starts appearing as soon as the question is sent.

The overall Chat mode experience is quite good. Compared to other chat tools, structured output is a clear plus.

For example, I wanted EvoX to help me design a performance benchmark test for my open-source project PaiCLI Agent.

The structured visualization panel is clear at a glance. Five sections (Metrics / Scenarios / Process / Scoring / Delivery Checklist) switch by clicking tabs, and the scoring model even includes an adjustable weight calculator. It's really much more comfortable than a typical text reply.

But Chat is just EvoX's basic capability. What I find most interesting is the Swarm Mode we'll talk about next.

03. Swarm Mode

I don't know if you've encountered this, but I've run into it more than once.

The context gets longer and longer, and earlier reasoning results start interfering with later judgments. After searching direction A and then direction B, by the time you're searching B, the key conclusions from A have already been diluted.

This leads to a situation where even with a Main Agent + Sub-Agent organizational structure, information is easily lost when the Sub-Agent reports back to the Main Agent at the end.

Smart you should know that ants are a very powerful social animal. Each ant only cares about what's right in front of it, and information is transmitted directly via pheromones, without any intermediary relaying it (I forget if this was taught in middle school textbooks).

The intelligence that emerges from the whole far exceeds that of the individual.

The intelligence of individual ants hasn't changed; what has changed is their way of organizing.

This organizational method is also applicable to Agents. EvoX's multi-Agent architecture is similar to the Swarm Mode.

For Agents, when facing complex tasks, there are usually two approaches.

The first is a single Agent, single thread. One AI does everything from start to finish in the same context. The context piles up longer and longer, earlier reasoning results interfere with later judgments, and the quality of the task degrades the further it goes.

The second is Main Agent + Sub-agent orchestration. The Main Agent breaks down the task, Sub-agents handle parts separately, and then hand the results back to the Main Agent for summarization.

It sounds like teamwork, but it ultimately still goes through a second round of paraphrasing and summarization by the Main Agent. Correct answers are easily lost during the aggregation phase because the Main Agent inevitably makes trade-offs when summarizing.

EvoX's Swarm Mode does three things differently.

Tasks are broken down more finely. Each Agent only handles one clearly bounded sub-problem, drastically reducing the complexity of a single task. An Agent doesn't need to remember ten things at once; it just needs to do the one thing it's responsible for well.

Contexts are isolated from each other. Each Agent only sees its own part and is not interfered with by the intermediate results of other Agents.

Results are merged deterministically. This is the biggest difference from traditional 'Main Agent + Sub-agent' orchestration.

In traditional orchestration, after a Sub-agent finishes, it hands the results back to the Main Agent, which then does another round of summarization and paraphrasing. This paraphrasing process itself loses information.

For example, the Main Agent might think a certain detail is unimportant and omit it, but that detail might be the key part of the correct answer.

In EvoX's Swarm Mode, Agents are only responsible for reasoning. The final output does not go through a second round of paraphrasing, meaning the original answer from each Agent is visible to the program.

Reasoning is reasoning, collection is collection; the responsibilities are clearly separated.

There's a set of data that can visually illustrate the effect. Using the same model, facing 563 logic, math, and physics problems, the accuracy rates of the three organizational methods differ greatly.

Single Agent, single thread only achieved 26%. Main Agent + Sub-agent orchestration improved, but information was still lost during the aggregation phase.

EvoX's Swarm Mode soared directly to 71%.

Same model, same problems. The accuracy differed by nearly twofold just because of the organizational method.

The model didn't change, the problems didn't change; the only thing that changed was 'how to divide the work among Agents'.

Swarm Mode also brings an additional benefit: token savings.

Stuffing 256 sub-tasks into one context requires roughly 49.34 million tokens. After Swarm division, each Agent only processes its own piece, totaling roughly 1.04 million tokens, a difference of about 48 times.

For users who frequently run long tasks, a 48x difference in token cost is not a small amount.

What tasks is Swarm Mode suitable for?

From my own experience, these directions are a natural fit:

Enter Code mode, select [Split Parallel Sessions], and enter the prompt.

Conduct a comprehensive code health review for the PaiCLI project, a terminal AI Agent tool implemented in Java (benchmarked against Claude Code), analyzing in parallel from the following 4 dimensions, with each dimension outputting an independent report:

  1. Security Audit: Scan whether the API Key handling process is secure, whether user input is validated, and whether there are injection risks in command execution.
  2. Performance Hotspots: Identify performance bottlenecks in the Agent reasoning loop, including context concatenation efficiency, tool call latency, and memory usage.
  3. Code Architecture: Evaluate the implementation quality of the ReAct and Plan-Execute dual modes, module coupling, and extensibility.
  4. Dependency Health: Check for outdated dependencies, known CVE vulnerabilities, and version conflicts.

Finally, aggregate into a comprehensive evaluation report, sorted by severity.

Click [Generate 5 Swarm Sessions].

You can see 5 Agents starting to work.

Ant colony pheromones and EvoX's Swarm have one thing in common: decentralization.

There is no 'commander-in-chief' centrally coordinating.

Clicking on [Tasks] allows you to see specific details.

The overall task complexity is very high.

Wait for the report.

The conclusions are very valuable for reference; I'll go upgrade and iterate right away.

04. Networked Self-Evolution

Swarm is a change in organizational method; networked self-evolution is a change in learning method. Putting these two things together is the biggest differentiator between EvoX and other Agent tools.

Let's talk about local evolution first.

EvoX saves used experiences locally.

For example, if it helped you organize a weekly report, the next time you do a similar task, it will reuse the previous experience. Your preferred format, phrasing style, and common data sources don't need to be taught again.

The more you use it, the more familiar the Agent becomes with your work habits. This is a gradual process; the effect becomes more pronounced over time.

But the more interesting part of EvoX is networked evolution.

EvoX is connected to the EvoMap experience network behind the scenes. You can think of EvoMap as a search engine for Agents, where a large amount of experience left by other Agents after completing tasks has accumulated.

EvoMap packages these experiences into two forms: Gene and Capsule.

A Gene is the smallest unit of experience, recording the successful execution path of a specific task. For example, a complete execution record of 'how to scrape a certain type of data with Python and clean it,' including what tools were used, what pitfalls were encountered along the way, and how they were ultimately resolved.

A Capsule is a packaged capability module that allows an Agent to quickly acquire a certain skill. A Capsule may contain multiple Genes, combined into a reusable capability package.

It's like stringing scattered experience points into a complete operational process.

Most Agents on the market that claim to get smarter the more you use them might actually never reach that 'smart' moment (no offense).

A new Agent, on its first day, already has experience available. It can inherit the execution experience accumulated by other Agents on the EvoMap network, directly reusing verified solutions when encountering similar tasks.

It's like a new employee being able to flip through all the operation manuals and experience summaries in the company's knowledge base on their first day, without having to figure everything out from scratch. The starting point is different. And as more Agents join EvoMap, the experience pool continues to expand.

EvoX already has a batch of high-quality Capsules from EvoMap built-in, ready to use out of the box. This is also the meaning behind the claim of being 'smart from day one.'

ending

After installing EvoX and running it for a day, here are a few honest impressions.

The idea behind Swarm Mode is eye-opening.

Task splitting, context isolation, and raw result aggregation—each step addresses a real pain point of current Agent tools.

Multi-model support is also a plus. DeepSeek V4 Flash, GLM-5.2, Qwen3.8 Max, GPT-5.6 Sol can be switched at will.

Different models excel in different scenarios, and the flexibility of mixing and matching them is very high. Picking one model alone might have its pros and cons, but EvoX does a great job of 'letting users choose the model that best fits their scenario.'

The vision of networked self-evolution is also excellent.

To be honest, EvoX just released its beta version. Some features are still being polished, and there is room for improvement in interaction details.

But the next step in Agent evolution is very likely to happen at the organizational level, enabling a group of AIs to divide work, inherit experience, and complete complex tasks together. EvoX has taken a concrete step in this direction.

It can be felt that behind EvoX is a very geeky young team. The product's foresight is very impressive, full of ideas, and has a great sense. From EvoMap to EvoX, they have carved out their own path and might become a dark horse among startups in the general Agent track, outside the big tech ecosystem.

EvoX is currently free to register, no invitation code needed. Interested friends, go to the official website to download and try it.

See you next time.