<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>AI Engineering on Arjun Sehajpal</title><link>https://arjunsehajpal.com/tags/ai-engineering/</link><description>Recent content in AI Engineering on Arjun Sehajpal</description><generator>Hugo</generator><language>en</language><lastBuildDate>Thu, 04 Jun 2026 16:29:55 +0530</lastBuildDate><atom:link href="https://arjunsehajpal.com/tags/ai-engineering/index.xml" rel="self" type="application/rss+xml"/><item><title>Context Engineering for Agents</title><link>https://arjunsehajpal.com/posts/07-context-engineering-for-agents/</link><pubDate>Thu, 04 Jun 2026 16:29:55 +0530</pubDate><guid>https://arjunsehajpal.com/posts/07-context-engineering-for-agents/</guid><description>&lt;p&gt;When the long context came into being, there was a lot of excitement around them. It threatened to disrupt the RAG&amp;rsquo;s status quo and pumped the hype for MCPs. Of what use is the carefully crafted RAG pipeline to fetch the best document when one can put every document in the context itself. The Frontier Model Labs latched onto the trend. &lt;a href="https://x.ai/news/grok-4-1-fast" class="external-link" target="_blank" rel="noopener"&gt;Grok 4.1 was launched with a 2 million context window&lt;/a&gt; which made the Gemini 1 million context window look rather ordinary. But these numbers say more about how much text the LLMs can process without breaking. There’s no guarantee that all the context present in such a large window would be actively reasoned. Chroma published a technical report explaining this, naming this phenomenon as &lt;a href="https://www.trychroma.com/research/context-rot" class="external-link" target="_blank" rel="noopener"&gt;Context Rot&lt;/a&gt;.&lt;/p&gt;</description></item><item><title>Spec-driven Development</title><link>https://arjunsehajpal.com/posts/08-spec-driven-development/</link><pubDate>Sun, 21 Dec 2025 01:19:02 +0530</pubDate><guid>https://arjunsehajpal.com/posts/08-spec-driven-development/</guid><description>&lt;p&gt;The emergence of AI coding agents has fundamentally transformed how the software is developed. They’re no longer a novelty, but a legitimate part of a developer’s toolkit. Though early coding assistants like GitHub Copilot were only able to generate small code snippets and perform completions, as LLMs improved and the context window grew larger, it became easier to generate software directly from natural language instructions. This gave rise to vibe-coding. &lt;/br&gt;
Vibe-coding is great for greenfield use cases, prototyping and creative exploration, but is not equally performant on the existing codebases and legacy systems. This is because vibe-coding treats software development as a series of isolated interactions. This unstructured chain of interactions where one describes a goal, prompts the LLM and repeats until the code “feels right”, has led to proliferation of anti-patterns and code smells.&lt;/br&gt;
To overcome this, we need a coherent process that reasserts the primacy of structured thinking and intentional design. One of the software industry’s more rigorous answers to fill this gap is Spec-driven development (SDD).&lt;/br&gt;&lt;/p&gt;</description></item></channel></rss>