<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Technical Writing on Viswa Balaji</title><link>https://viswabalaji.github.io/technical/</link><description>Recent content in Technical Writing on Viswa Balaji</description><generator>Hugo -- gohugo.io</generator><language>en-us</language><lastBuildDate>Wed, 28 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://viswabalaji.github.io/technical/index.xml" rel="self" type="application/rss+xml"/><item><title>Microsoft Copilot Harness in Foundry Agent Service</title><link>https://viswabalaji.github.io/technical/microsoft-copilot-harness-foundry/</link><pubDate>Wed, 28 Oct 2026 00:00:00 +0000</pubDate><guid>https://viswabalaji.github.io/technical/microsoft-copilot-harness-foundry/</guid><description>&lt;p&gt;This week we launched the &lt;strong&gt;Microsoft Copilot harness in public preview in Foundry Agent Service&lt;/strong&gt;. You define the agent declaratively with instructions, tools, skills, and guardrails, and Microsoft manages the execution system that runs it.&lt;/p&gt;
&lt;p&gt;&lt;img src="https://viswabalaji.github.io/images/technical/microsoft-copilot-harness/microsoft-copilot-harness-high-level.png" alt="Governed reasoning and isolated tool execution: prompt and Routine inputs, a Foundry model and Microsoft Copilot harness, a named skill, governed tools, Code Interpreter, and a separate durable conversation store."&gt;&lt;/p&gt;
&lt;p&gt;The harness is the same multi-model execution layer Microsoft has standardized on across its Copilot experiences, and &lt;a href="https://www.linkedin.com/posts/satyanadella_were-redefining-the-sdlc-for-the-agentic-ugcPost-7521256751313375232-ObpB/"&gt;Satya announced it&lt;/a&gt; at GitHub Universe as part of a broader push to redefine the software development lifecycle for the agentic era. &lt;a href="https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/"&gt;Tina Schuchman and Vladimir Fedorov&amp;rsquo;s post&lt;/a&gt; explains why a shared harness matters: models change quickly, and every improvement to the harness now reaches agents built on Foundry too. Jeff Hollan and I wrote the &lt;a href="https://devblogs.microsoft.com/foundry/microsoft-copilot-harness-in-foundry/"&gt;launch post on the Foundry blog&lt;/a&gt;, which &lt;a href="https://www.linkedin.com/feed/update/urn:li:activity:7521266220462018560/"&gt;Satya also shared&lt;/a&gt;. Here I go deeper with a fuller example.&lt;/p&gt;</description></item><item><title>GEPA, Pareto Frontiers, and the Art of Staying Adaptable</title><link>https://viswabalaji.github.io/technical/genetic-pareto-frontier-adaptability/</link><pubDate>Sat, 18 Apr 2026 00:00:00 +0000</pubDate><guid>https://viswabalaji.github.io/technical/genetic-pareto-frontier-adaptability/</guid><description>&lt;p&gt;I have been spending a lot of time in the Agentic AI space, and one pattern keeps repeating across serious systems. Model capability keeps increasing, and memory, retrieval, and context-handling architectures keep improving, but production behavior still depends on a third layer: how well we configure the agent with the right prompts, tools, retrieval resources, and execution policy for the actual query distribution seen at runtime.&lt;/p&gt;
&lt;p&gt;That quickly becomes a design question. Do we want an agent that is a master of one, or an agent that is a jack of all trades? A specialist can dominate on precision, policy compliance, and bounded latency in a narrow operating region. A generalist can dominate on coverage and adaptability when requirements move. The mistake is not picking one or the other; the mistake is assuming there is one universally superior choice independent of objectives and constraints.&lt;/p&gt;</description></item><item><title>Daytona Compute + PyAI Conference</title><link>https://viswabalaji.github.io/technical/daytona-pyai-conference/</link><pubDate>Thu, 12 Mar 2026 00:00:00 +0000</pubDate><guid>https://viswabalaji.github.io/technical/daytona-pyai-conference/</guid><description>&lt;p&gt;Last week at Daytona Compute and PyAI conferences, the same architectural pattern dominated: agent-in-sandbox. These agents run in full OS environments with forking, checkpointing, and multi-day execution. &lt;a href="https://www.daytona.io/"&gt;Daytona&lt;/a&gt; demonstrated 50,000 concurrent sandboxes in 2 minutes 13 seconds. Multiple companies showed agents running for days.&lt;/p&gt;
&lt;p&gt;Five patterns from the floor:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Sandbox architecture choices&lt;/strong&gt;: agent-in-sandbox versus sandbox-as-tool patterns&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Progressive tool disclosure&lt;/strong&gt;: using knowledge graphs to manage complexity, not just for skills but for code understanding&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Browser vs computer use&lt;/strong&gt;: when each approach makes sense&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;CPU bottleneck&lt;/strong&gt;: why there&amp;rsquo;s a high spike in CPU demand again&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Parallel execution&lt;/strong&gt;: multiple long-running agents per user&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="sandbox-patterns-agent-in-vs-tool"&gt;Sandbox Patterns: Agent-In vs Tool&lt;/h2&gt;
&lt;p&gt;Two distinct patterns emerged. In &lt;a href="https://blog.langchain.com/the-two-patterns-by-which-agents-connect-sandboxes/"&gt;agent-in-sandbox&lt;/a&gt;, the agent runs inside a full OS environment with persistent state, installing packages and maintaining processes. The sandbox becomes the workspace. In sandbox-as-tool, the agent operates outside and calls the sandbox for specific stateless execution via APIs from providers like E2B, Modal, Daytona, or Runloop.&lt;/p&gt;</description></item><item><title>Teaching Agents Intuition: Programmatic Tool Chain Caching</title><link>https://viswabalaji.github.io/technical/programmatic-tool-chain-caching/</link><pubDate>Tue, 10 Mar 2026 00:00:00 +0000</pubDate><guid>https://viswabalaji.github.io/technical/programmatic-tool-chain-caching/</guid><description>&lt;h2 id="the-recomputation-problem"&gt;The Recomputation Problem&lt;/h2&gt;
&lt;p&gt;Ask an agent to analyze customer churn. It explores tool compositions, tries different approaches, settles on a working sequence. Thirty seconds later, it succeeds. Ask it again tomorrow. It starts from scratch. Same exploration, same trial and error, same thirty seconds. No memory of what worked. No learning.&lt;/p&gt;
&lt;p&gt;Software engineers learned decades ago to cache expensive computations and reuse proven solutions. AI agents are still figuring out the same problems repeatedly.&lt;/p&gt;</description></item><item><title>Reflections on Amazon Bedrock Inference Team</title><link>https://viswabalaji.github.io/technical/reflections-amazon-bedrock-inference-team/</link><pubDate>Tue, 02 Sep 2025 00:00:00 +0000</pubDate><guid>https://viswabalaji.github.io/technical/reflections-amazon-bedrock-inference-team/</guid><description>&lt;p&gt;In September 2023, I moved to a new team that was being setup. I didn&amp;rsquo;t know who my manager was going to be or the exact nature of this project. Later I&amp;rsquo;d find out that we&amp;rsquo;d be integrating Anthropic&amp;rsquo;s Claude models into AWS infrastructure as part of the new &lt;a href="https://press.aboutamazon.com/2023/9/amazon-and-anthropic-announce-strategic-collaboration-to-advance-generative-ai"&gt;Amazon-Anthropic partnership&lt;/a&gt;. Two years later, I&amp;rsquo;ve helped scale one of the fastest-growing AI services in AWS history.&lt;/p&gt;
&lt;p&gt;&lt;a href="https://aws.amazon.com/bedrock/"&gt;Amazon Bedrock&lt;/a&gt; processes millions of inference requests daily for Fortune 500 companies and startups alike. High availability, low latency, incredible rate of feature delivery (pre coding assistant era) required careful thought and management in a competitive market. In my first year, I helped launch the &lt;a href="https://aws.amazon.com/blogs/machine-learning/fine-tune-anthropics-claude-3-haiku-in-amazon-bedrock-to-boost-model-accuracy-and-quality/"&gt;model fine-tuning feature for Anthropic&amp;rsquo;s Claude models&lt;/a&gt; from private beta through GA. In my second year, I worked on Anthropic model inference services, maturing them into production-grade systems. I&amp;rsquo;ve since moved to AWS&amp;rsquo;s Agentic AI organization. My core motivation for writing this document stems from the quote &lt;em&gt;&amp;ldquo;You don&amp;rsquo;t learn from experiences, you learn by reflecting on your experiences&amp;rdquo;&lt;/em&gt;—a foresight my first manager in the industry instilled in me. My core reason for sharing these reflections broadly was inspired by the quote &lt;em&gt;&amp;ldquo;We make our living by what we get, but we make our life by what we give&amp;rdquo;&lt;/em&gt;.&lt;/p&gt;</description></item></channel></rss>