<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Parth Gupta</title><description>I build agent harnesses: the loop, tools, context and evals around the model. Most of the work is the infra that keeps autonomous loops from doing something stupid.</description><link>https://parthg.dev/</link><item><title>Stop fixing your agent in the prompt</title><link>https://parthg.dev/posts/stop-fixing-your-agent-in-the-prompt/</link><guid isPermaLink="true">https://parthg.dev/posts/stop-fixing-your-agent-in-the-prompt/</guid><description>A prompt patch after one bad run bets twice: that the failure will repeat, and that the prompt caused it. Why both bets usually lose, and what to do instead.</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Most prompt cache misses are self-inflicted</title><link>https://parthg.dev/posts/most-prompt-cache-misses-are-self-inflicted/</link><guid isPermaLink="true">https://parthg.dev/posts/most-prompt-cache-misses-are-self-inflicted/</guid><description>A cache hit rate that looks healthy can hide the requests where your agent harness rewrote its own prefix. How to find them, and why the provider is rarely the one to blame.</description><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate></item><item><title>The browser-shaped hole in ACP</title><link>https://parthg.dev/posts/the-browser-shaped-hole-in-acp/</link><guid isPermaLink="true">https://parthg.dev/posts/the-browser-shaped-hole-in-acp/</guid><description>ACP got the primitives right, but the spec stops at stdio. What gets reinvented when teams build to the browser, and what closing the gap would require.</description><pubDate>Sat, 09 May 2026 00:00:00 GMT</pubDate></item></channel></rss>