跪拜 Guibai
← All articles
Agent · AI Programming

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

By 沉默王二 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most multi-Agent tools still funnel results through a master Agent that summarizes, introducing a lossy bottleneck. EvoX's deterministic merge and 48x token reduction show that organizational architecture, not just model capability, determines whether an Agent system can handle real engineering tasks without hallucinating or dropping critical details.

Summary

EvoX is a free beta desktop Agent for Mac and Windows that supports DeepSeek V4 Flash, GLM-5.2, Qwen3.8 Max, GPT-5.6 Sol, and local models. Its core differentiator is a Swarm Mode that breaks complex tasks into isolated sub-problems handled by separate Agents, then deterministically merges the raw outputs. No master Agent paraphrases or summarizes the results, which eliminates a common failure point where critical details get dropped during aggregation. On a benchmark of 563 logic, math, and physics problems, the same model scored 26% accuracy in a single-threaded run and 71% under Swarm Mode, while token consumption dropped roughly 48x compared to stuffing all sub-tasks into one context.

A networked self-evolution system connects EvoX to EvoMap, a shared experience network where completed task paths are stored as Genes and reusable skill bundles as Capsules. A fresh install inherits verified execution patterns from other Agents on the network, so it starts with domain knowledge instead of a blank slate. Local experience also accumulates per user, adapting to individual formatting and workflow preferences over time.

The product splits into three workspaces: Chat for structured Q&A with auto-generated tables and charts, Cowork for office automation and long-running task flows, and Code for reading projects, modifying code, and running tests. Registration is open with no invite code required.

Takeaways
Swarm Mode splits a task into isolated sub-problems, each handled by a dedicated Agent with its own context, and merges raw outputs without a summarizing master Agent.
On 563 logic, math, and physics problems, the same model scored 26% in single-threaded mode and 71% under Swarm Mode.
Token consumption for 256 sub-tasks dropped from roughly 49.34 million tokens in a single context to about 1.04 million tokens under Swarm Mode, a ~48x reduction.
EvoX connects to EvoMap, a shared network where Agents publish successful execution paths (Genes) and reusable skill bundles (Capsules) that other Agents can inherit.
Built-in models include DeepSeek V4, Kimi K3, Opus 4.8, and GPT-5.6 Sol; users can also configure their own API keys for DeepSeek V4 Flash, GLM-5.2, Qwen3.8 Max, and local models.
The product offers three modes: Chat (structured Q&A), Cowork (office automation and long-running tasks), and Code (project reading, code modification, testing).
Registration is free with no invitation code; the beta launched August 9 for Mac and Windows.
Conclusions

Swarm Mode's 71% accuracy on a reasoning benchmark is a concrete data point that organizational architecture can nearly triple a model's effective performance without changing the model itself.

The 48x token reduction is a direct cost argument for context-isolated multi-Agent architectures over monolithic long-context approaches, especially for teams running repeated complex evaluations.

EvoMap's shared experience network inverts the typical Agent learning curve: instead of starting dumb and improving locally over weeks, a new Agent inherits verified execution patterns from day one, which changes the onboarding economics for teams adopting Agent tooling.

The deterministic merge approach sidesteps a subtle failure mode in Agent orchestration where a master Agent's summarization step silently drops the one detail that was the correct answer, a problem that is hard to debug because the omission looks plausible.

Concepts & terms
Swarm Mode
A decentralized multi-Agent architecture where tasks are split into isolated sub-problems, each Agent works independently with its own context, and raw outputs are merged deterministically without a summarizing master Agent.
Gene
In EvoMap's experience network, a Gene is the smallest unit of recorded experience, capturing the full execution path of a specific task including tools used, pitfalls encountered, and the final solution.
Capsule
A packaged, reusable capability module on EvoMap that bundles multiple Genes into a complete skill set, allowing an Agent to quickly acquire a new competency without training from scratch.
Context Pollution
The degradation of reasoning quality that occurs when an Agent's context window accumulates too much information, causing earlier conclusions to interfere with later judgments and diluting key findings.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗