<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Production on Programmer.ie: Modern AI programming</title>
    <link>http://programmer.ie/tags/production/</link>
    <description>Recent content in Production on Programmer.ie: Modern AI programming</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Sun, 09 Aug 2026 16:09:00 +0100</lastBuildDate>
    <atom:link href="http://programmer.ie/tags/production/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Build a Production AI Agent From First Principles: The Complete Reference Architecture</title>
      <link>http://programmer.ie/post/advanced-agents-from-first-principles-44/</link>
      <pubDate>Sun, 09 Aug 2026 16:09:00 +0100</pubDate>
      <guid>http://programmer.ie/post/advanced-agents-from-first-principles-44/</guid>
      <description>&lt;h1 id=&#34;build-a-production-ai-agent-from-first-principles-the-complete-reference-architecture&#34;&gt;Build a Production AI Agent From First Principles: The Complete Reference Architecture&lt;/h1&gt;&#xA;&lt;p&gt;We have spent this series adding mechanisms only when a specific failure demanded them.&lt;/p&gt;&#xA;&lt;p&gt;We started with a model call.&lt;/p&gt;&#xA;&lt;p&gt;Then we added candidate generation, critique, planning, tool use, memory, search and verification.&lt;/p&gt;&#xA;&lt;p&gt;Then the system stopped looking like a clever prompt.&lt;/p&gt;&#xA;&lt;p&gt;It started looking like software.&lt;/p&gt;&#xA;&lt;p&gt;Then distributed systems problems arrived:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;duplicate work,&lt;/li&gt;&#xA;&lt;li&gt;retries,&lt;/li&gt;&#xA;&lt;li&gt;leases,&lt;/li&gt;&#xA;&lt;li&gt;fencing,&lt;/li&gt;&#xA;&lt;li&gt;backpressure,&lt;/li&gt;&#xA;&lt;li&gt;dependency failure,&lt;/li&gt;&#xA;&lt;li&gt;behavioral drift,&lt;/li&gt;&#xA;&lt;li&gt;release compatibility,&lt;/li&gt;&#xA;&lt;li&gt;replay,&lt;/li&gt;&#xA;&lt;li&gt;incident forensics,&lt;/li&gt;&#xA;&lt;li&gt;SLOs,&lt;/li&gt;&#xA;&lt;li&gt;authority,&lt;/li&gt;&#xA;&lt;li&gt;competence,&lt;/li&gt;&#xA;&lt;li&gt;capability acquisition,&lt;/li&gt;&#xA;&lt;li&gt;placement,&lt;/li&gt;&#xA;&lt;li&gt;handoff,&lt;/li&gt;&#xA;&lt;li&gt;stale state,&lt;/li&gt;&#xA;&lt;li&gt;stale intent,&lt;/li&gt;&#xA;&lt;li&gt;commitments,&lt;/li&gt;&#xA;&lt;li&gt;durable workflows,&lt;/li&gt;&#xA;&lt;li&gt;transaction recovery,&lt;/li&gt;&#xA;&lt;li&gt;trust boundaries,&lt;/li&gt;&#xA;&lt;li&gt;multi-agent coordination,&lt;/li&gt;&#xA;&lt;li&gt;and finally an explicit control plane.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;At this point the architecture is complete enough that adding another isolated mechanism would make the series worse rather than better.&lt;/p&gt;</description>
    </item>
    <item>
      <title>Advanced Agents From First Principles 24: How Do You Release Agent Behavior Safely? Add Behavioral Contracts, Compatibility Checks and Promotion Gates</title>
      <link>http://programmer.ie/post/advanced-agents-from-first-principles-24/</link>
      <pubDate>Sun, 09 Aug 2026 12:09:00 +0100</pubDate>
      <guid>http://programmer.ie/post/advanced-agents-from-first-principles-24/</guid>
      <description>&lt;p&gt;A model change can pass health checks and still break your agent.&lt;/p&gt;&#xA;&lt;p&gt;A prompt change can preserve output schema and still change which tools the agent chooses.&lt;/p&gt;&#xA;&lt;p&gt;A router change can improve average cost and silently starve a specialist that is essential for one high-risk workload.&lt;/p&gt;&#xA;&lt;p&gt;A memory migration can keep every row intact and still change which memories are retrieved.&lt;/p&gt;&#xA;&lt;p&gt;A verifier upgrade can make every dashboard greener while making the system less trustworthy.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
