<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Agentic-Infrastructure on Jeff Geiser</title>
    <link>https://jeffgeiser.dev/tags/agentic-infrastructure/</link>
    <description>Recent content in Agentic-Infrastructure on Jeff Geiser</description>
    <generator>Hugo</generator>
    <language>en-US</language>
    <lastBuildDate>Wed, 26 Aug 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://jeffgeiser.dev/tags/agentic-infrastructure/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>The Layer Below MCP</title>
      <link>https://jeffgeiser.dev/posts/the-layer-below-mcp/</link>
      <pubDate>Wed, 26 Aug 2026 00:00:00 +0000</pubDate>
      <guid>https://jeffgeiser.dev/posts/the-layer-below-mcp/</guid>
      <description>&lt;p&gt;&lt;strong&gt;Connection isn&amp;rsquo;t the hard part anymore. Reasoning about resources is.&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;em&gt;By Jeff Geiser — April 2026&lt;/em&gt;&lt;/p&gt;&#xA;&lt;hr&gt;&#xA;&lt;p&gt;Today, most agents run on a single machine or call a public API. Routing is trivial — there&amp;rsquo;s one option. The question &amp;ldquo;which GPU should this go to&amp;rdquo; doesn&amp;rsquo;t come up because there&amp;rsquo;s nothing to pick between.&lt;/p&gt;&#xA;&lt;p&gt;That changes the moment an agent has a fleet under it. One GPU is cool and idle, with plenty of free VRAM. Another is thermally throttled, busy, near OOM. The question of where to send the work becomes real but the agent isn&amp;rsquo;t the one answering it. A layer underneath is.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
