<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Pi on Han's XYZ</title><link>https://han8931.github.io/tags/pi/</link><description>Recent content in Pi on Han's XYZ</description><generator>Hugo</generator><language>en</language><managingEditor>tabularasa8931@gmail.com (Han)</managingEditor><webMaster>tabularasa8931@gmail.com (Han)</webMaster><copyright>This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.</copyright><lastBuildDate>Sun, 23 Aug 2026 21:57:42 +0900</lastBuildDate><atom:link href="https://han8931.github.io/tags/pi/index.xml" rel="self" type="application/rss+xml"/><item><title>Deep Dive into Agents</title><link>https://han8931.github.io/agent-harness/</link><pubDate>Sun, 23 Aug 2026 00:00:00 +0000</pubDate><author>tabularasa8931@gmail.com (Han)</author><guid>https://han8931.github.io/agent-harness/</guid><description>&lt;p&gt;The word &lt;em&gt;agent&lt;/em&gt; gets used very loosely — sometimes it means a chatbot with a tool attached, sometimes a long-running autonomous system. This talk uses a narrower engineering definition and then spends its time on the part nobody puts on a leaderboard: &lt;strong&gt;the harness&lt;/strong&gt;, the software wrapped around the model call.&lt;/p&gt;
&lt;p&gt;The claim in one line: two agents running the same model weights can differ by seven to ten benchmark points, and the difference is entirely the harness. If that holds, choosing an agent is not a model-selection problem — it&amp;rsquo;s a systems-engineering one.&lt;/p&gt;</description></item></channel></rss>