<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Prompt-Injection on tech.mytrapp.com</title><link>https://tech.mytrapp.com/tags/prompt-injection/</link><description>Recent content in Prompt-Injection on tech.mytrapp.com</description><generator>Hugo</generator><language>en-us</language><copyright>David Trapp</copyright><lastBuildDate>Sun, 25 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://tech.mytrapp.com/tags/prompt-injection/index.xml" rel="self" type="application/rss+xml"/><item><title>Agent Harness Engineering: Twenty-Four Practices from an Instrumented Case Study</title><link>https://tech.mytrapp.com/posts/harness-engineering-agent-practices/</link><pubDate>Sun, 25 Oct 2026 00:00:00 +0000</pubDate><guid>https://tech.mytrapp.com/posts/harness-engineering-agent-practices/</guid><description>One single-file agent harness and twenty-four practices that survived it, each traced to the failure that paid for it: a 4,116-token fixed overhead rendered from the tree and held to a ceiling, output caps derived from the window the endpoint declares, confirmation tiers that read parsed effect instead of spelling, and an injection defence I built after watching a run execute four numbered steps it had only been asked to read.</description></item></channel></rss>