Splitting planning from clicking made my desktop agent much faster: frontier model plans, small model acts
Summary
作者分享了将桌面智能体的规划与执行分离的架构：由 Claude Code 等前沿模型批量规划步骤，小型快速模型 Jev 负责执行点击与操作，普通步骤跳过模型，从而显著提升速度与降低成本，测试中 12 个随机按钮达成 47/48、约 1.2 秒每次点击。代码已开源。
Similar Articles
@GoSailGlobal: Practical data on multi-agent AI collaboration: Use Opus 4.8 for planning, Deepseek/Gemma for execution — 10x cost reduction, 2x speed improvement. The secret is not using the most expensive model, but having cheap models do the heavy lifting and expensive models only make decisions. This is the same as company management: the CEO shouldn't write code, and interns shouldn't set strategy. A…
A practical sharing on multi-agent AI collaboration, proposing a hierarchical strategy using Opus 4.8 for planning and Deepseek/Gemma for execution, achieving a 10x cost reduction and 2x speed improvement, with open-source implementation.
@AI_Community_9: Unsloth Desktop takes the experience of Qwen3.8 27B to new heights. Out-of-the-box + cross-platform, the core is optimizing VRAM for fine-tuning and inference to the extreme (VRAM saved by 70%, speed increased by 2 times). It packs 'training, inference, and Agent toolchain' into a free, open-source desktop application, …
Unsloth Desktop is a free, open-source desktop application that integrates training, inference, and Agent toolchain through extreme optimization of VRAM and speed, significantly improving the user experience of AI models like Qwen3.8 27B.
When working on agent workflows or autonomous agent loops, a major bottleneck affecting experience is: every small branch decision, tool routing, task completion check, or safety interception, if all handled by expensive reasoning LLMs, not only brings seconds of delay per call, but token costs also...
This article introduces the open-source list awesome-jev-projects, which decouples System 1 fast decision models from System 2 reasoning LLMs to optimize AI agent workflows, reduce latency and costs, and provide code-level references.
@yibie: Recommend this project. shadcn (creator of shadcn/ui) built an Agent Skill — let the most expensive model handle auditing and planning, and let the cheap model write code. The idea itself isn't new, but shadcn turned it into an installable, orchestrated system with an execution closed loop. Every design decision is practical...
shadcn released the Agent Skill project improve, which has high-cost models perform code auditing and planning while low-cost models execute, forming an installable, orchestrated system with an execution closed loop.
@QingQ77: An operational Agent runtime built for llama.cpp local inference models, allowing local models to execute real-world tasks like browser, file, and Shell operations like a desktop operator https://github.com/AtomicBot-ai/atomic-agent… Atom…
Atomic-Agent is a desktop operation Agent designed for llama.cpp local inference models, optimizing the runtime architecture to enable small local models to reliably execute multi-step desktop tasks.