<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Technology on Yosgi Blog</title><link>https://siqi-liu.com/en/categories/technology/</link><description>Recent content in Technology on Yosgi Blog</description><generator>Hugo</generator><language>en</language><lastBuildDate>Fri, 12 Jun 2026 04:28:41 +0000</lastBuildDate><atom:link href="https://siqi-liu.com/en/categories/technology/index.xml" rel="self" type="application/rss+xml"/><item><title>Your Agent Is Not Inconsistent Because It Is Dumb. It Just Has Too Many Tool Paths.</title><link>https://siqi-liu.com/en/post/your-agent-is-not-inconsistent-because-it-is-dumb-it-just-has-too-many-tool-paths/</link><pubDate>Thu, 07 May 2026 00:00:00 +0000</pubDate><guid>https://siqi-liu.com/en/post/your-agent-is-not-inconsistent-because-it-is-dumb-it-just-has-too-many-tool-paths/</guid><description>Multi-tool agents become unreliable when the same task takes different tool paths each time. This article explains why Skills should preserve validated execution paths, not one-off results, and why those Skills need to evolve over time.</description></item><item><title>Tool for Control, Code for Analysis(3)</title><link>https://siqi-liu.com/en/post/tool-for-control-code-for-analysis3/</link><pubDate>Wed, 28 Jan 2026 00:00:00 +0000</pubDate><guid>https://siqi-liu.com/en/post/tool-for-control-code-for-analysis3/</guid><description>In large-scale Agent systems, pure Tool-based approaches tend to collapse under context pressure, while pure Code-based approaches introduce excessive latency. Based on real-world experiments, this article introduces a dual-path execution model that uses a “Context Off-Ramp” to switch between Tool and Code execution.</description></item><item><title>Why Tool and Code Fail Differently in Agent Systems（2）</title><link>https://siqi-liu.com/en/post/why-tool-and-code-fail-differently-in-agent-systems2/</link><pubDate>Sat, 27 Dec 2025 00:00:00 +0000</pubDate><guid>https://siqi-liu.com/en/post/why-tool-and-code-fail-differently-in-agent-systems2/</guid><description>Tool and Code do not just differ in expressiveness; they also fail at different times. This article explains why multi-round Agent systems amplify that difference.</description></item><item><title>An MCP-as-Code Refactor, and Why It Did Not Work the Way I Expected</title><link>https://siqi-liu.com/en/post/an-mcp-as-code-refactor-and-why-it-did-not-work-the-way-i-expected1/</link><pubDate>Thu, 18 Dec 2025 00:00:00 +0000</pubDate><guid>https://siqi-liu.com/en/post/an-mcp-as-code-refactor-and-why-it-did-not-work-the-way-i-expected1/</guid><description>A real-world MCP-as-Code refactor on where Code helps, where Tool still fits better, and why Agent runtime needs both.</description></item><item><title>High-Frequency Synchronization Architecture Between React State and a 3D Engine</title><link>https://siqi-liu.com/en/post/high-frequency-synchronization-architecture-between-react-state-and-a-3d-engine/</link><pubDate>Wed, 01 Oct 2025 00:00:00 +0000</pubDate><guid>https://siqi-liu.com/en/post/high-frequency-synchronization-architecture-between-react-state-and-a-3d-engine/</guid><description>A synchronization paradigm for massive real-time data: a middle layer isolates high-frequency data sources, and React consumes only the linear projection of the visible viewport.</description></item><item><title>Early Experiences with Building MCP Tools</title><link>https://siqi-liu.com/en/post/early-experiences-with-building-mcp-tools/</link><pubDate>Tue, 01 Jul 2025 00:00:00 +0000</pubDate><guid>https://siqi-liu.com/en/post/early-experiences-with-building-mcp-tools/</guid><description>How Cut 6,000 Tokens Down to 500</description></item></channel></rss>