<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title><![CDATA[Jigar Joshi — Notes]]></title>
    <link>https://jigarjoshi.in/notes</link>
    <description><![CDATA[Visual, carousel-style breakdowns of agentic AI concepts and patterns.]]></description>
    <language>en</language>
    <lastBuildDate>Mon, 15 Jun 2026 00:00:00 GMT</lastBuildDate>
    <atom:link href="https://jigarjoshi.in/notes/feed.xml" rel="self" type="application/rss+xml" />
    <item>
      <title><![CDATA[Agentic AI content quality: 5 agents, one pipeline.]]></title>
      <link>https://jigarjoshi.in/notes/agentic-ai-content-quality-system</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/agentic-ai-content-quality-system</guid>
      <pubDate>Mon, 15 Jun 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Production]]></category>
      <description><![CDATA[Most teams ask one model to evaluate and rewrite the same draft. That is how hallucinated scores and polished-but-wrong content slip through. This is the five-agent pipeline I use for GEO + SEO content QA: Orchestrator on Sonnet, read-only CORE and EEAT evaluators, Schema validator on Haiku, Rewrite on Opus only for failing sections. Input guardrails abort bad URLs before any token spend. Output guardrails recheck scores, fixes, citations, and JSON-LD before delivery. Three memory layers (CAG, RAG, session) keep context lean.]]></description>
    </item>
    <item>
      <title><![CDATA[Stop paying frontier prices for classification.]]></title>
      <link>https://jigarjoshi.in/notes/stop-paying-frontier-prices-model-routing</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/stop-paying-frontier-prices-model-routing</guid>
      <pubDate>Wed, 03 Jun 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Architecture]]></category>
      <description><![CDATA[Using one frontier model for every agent step is the most expensive mistake I still see on client bills. Classification belongs on Haiku or Flash. Tool calling belongs on Sonnet. Planning belongs on Opus. Embeddings are a different product entirely. The pattern that fixes the bill: a router agent on a cheap model classifies the task, then hands execution to the right tier. One visual map of all four tiers, the June 2026 model picks, and the bill audit that tells you whether you have a routing problem.]]></description>
    </item>
    <item>
      <title><![CDATA[Sequential or parallel? Draw the flow.]]></title>
      <link>https://jigarjoshi.in/notes/agentic-ai-sequential-vs-parallel-workflow</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/agentic-ai-sequential-vs-parallel-workflow</guid>
      <pubDate>Wed, 03 Jun 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Multi-Agent]]></category>
      <description><![CDATA[Most multi-agent systems run sequential when parallel would be faster, or parallel when correctness needs sequence. Sequential is for when the next agent needs the previous one's output, and it costs about 3x latency, the price of correctness. Parallel is for independent slices of the same task and can save up to 70% latency. The test before you code: if agents depend on each other, go sequential; if there is no dependency, go parallel.]]></description>
    </item>
    <item>
      <title><![CDATA[Wrong memory. Dead agent.]]></title>
      <link>https://jigarjoshi.in/notes/wrong-memory-dead-agent</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/wrong-memory-dead-agent</guid>
      <pubDate>Fri, 29 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Architecture]]></category>
      <description><![CDATA[Wrong memory choice kills more agents in production than wrong model choice. Context window, RAG, CAG, and a database each solve a different job. Use one for everything and you get slow retrieval, hallucinated answers, runaway token costs, and agents that forget the moment a session ends. A one-glance decision rule for matching the memory type to the problem.]]></description>
    </item>
    <item>
      <title><![CDATA[Your agents aren't broken. Your tools are.]]></title>
      <link>https://jigarjoshi.in/notes/agents-arent-broken-tools-are</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/agents-arent-broken-tools-are</guid>
      <pubDate>Mon, 25 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Tooling]]></category>
      <description><![CDATA[Most teams debug the agent when the real fault is the tool. An AI tool is a contract the model has to trust: atomic (one verb), honest on failure (semantic errors, never null), and typed and token-efficient (a schema, never raw DB rows). Three questions to ask before you build one. A one-glance visual companion to the long-form post on the tool contract.]]></description>
    </item>
    <item>
      <title><![CDATA[Agentic AI has an execution problem.]]></title>
      <link>https://jigarjoshi.in/notes/webmcp-agent-ready-websites</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/webmcp-agent-ready-websites</guid>
      <pubDate>Thu, 21 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[MCP]]></category>
      <description><![CDATA[Most AI agents still figure out how a website works by parsing the DOM, running OCR on screenshots, and hoping the click lands in the right place. Google previewed WebMCP at I/O 2026 to fix that: websites expose structured tools directly through a new browser API, agents call them by name. No more guessing. Agent-ready websites are coming.]]></description>
    </item>
    <item>
      <title><![CDATA[Your agent has no memory. That is the problem.]]></title>
      <link>https://jigarjoshi.in/notes/agent-memory-three-types</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/agent-memory-three-types</guid>
      <pubDate>Mon, 18 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Architecture]]></category>
      <description><![CDATA[Most AI agents forget the moment a session ends. Three memory layers fix that for good: implicit (baked into weights), explicit (RAG and vector DBs), and agentic (state that survives across sessions). A one-glance visual companion to the long-form post on the three paradigms of LLM memory.]]></description>
    </item>
    <item>
      <title><![CDATA[You have 3 tools. Are you using them correctly?]]></title>
      <link>https://jigarjoshi.in/notes/three-ai-tools-one-workflow</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/three-ai-tools-one-workflow</guid>
      <pubDate>Fri, 15 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Tooling]]></category>
      <description><![CDATA[Cursor, Claude Code, and Copilot each cover a different layer of the coding workflow — speed, depth, and quality. A 4-slide breakdown of who does what, when to hand off, and the one config file each tool needs so the boundaries hold under real production pressure.]]></description>
    </item>
    <item>
      <title><![CDATA[Your agent called the wrong tool.]]></title>
      <link>https://jigarjoshi.in/notes/agent-called-wrong-tool-fix-the-description</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/agent-called-wrong-tool-fix-the-description</guid>
      <pubDate>Wed, 13 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Tool Design]]></category>
      <description><![CDATA[When an agent picks the wrong tool, the agent is rarely the problem — the tool registry is. Three rules that fix tool selection in production: name tools precisely, describe when each one should be used, and load only what the task needs.]]></description>
    </item>
    <item>
      <title><![CDATA[Copilot vs Claude Code]]></title>
      <link>https://jigarjoshi.in/notes/copilot-vs-claude-code</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/copilot-vs-claude-code</guid>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Tooling]]></category>
      <description><![CDATA[A 4-slide breakdown of when to reach for Copilot vs Claude Code — by scope, context, and where the code is going. The cheat-sheet most teams need before they pick a default coding assistant.]]></description>
    </item>
    <item>
      <title><![CDATA[You merged it. Then it failed.]]></title>
      <link>https://jigarjoshi.in/notes/you-merged-it-then-it-failed-pr-review-pipeline</link>
      <guid isPermaLink="true">https://jigarjoshi.in/notes/you-merged-it-then-it-failed-pr-review-pipeline</guid>
      <pubDate>Thu, 07 May 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Tooling]]></category>
      <description><![CDATA[A single-image breakdown of the 3-stage automated review pipeline that runs before you merge: code review, security scan, and test generation — all triggered from a single PR open. Ship faster. Break less.]]></description>
    </item>
  </channel>
</rss>