<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Anshita Bhasin on Medium]]></title>
        <description><![CDATA[Stories by Anshita Bhasin on Medium]]></description>
        <link>https://medium.com/@anshita.bhasin?source=rss-cf4a02363eb1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ZhcUuMOy4d8hA45QgKmMHg.png</url>
            <title>Stories by Anshita Bhasin on Medium</title>
            <link>https://medium.com/@anshita.bhasin?source=rss-cf4a02363eb1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:42:55 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anshita.bhasin/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Bug That Hid in 70% of the Web for 10+ Years, and How AI Found It]]></title>
            <link>https://medium.com/@anshita.bhasin/the-bug-that-hid-in-70-of-the-web-for-10-years-and-how-ai-found-it-a6e13716dba1?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a6e13716dba1</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[bugs]]></category>
            <category><![CDATA[php]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 20:02:53 GMT</pubDate>
            <atom:updated>2026-10-22T20:02:53.922Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vBtmKfC7HM8SFWWU0cYIkQ.png" /></figure><p>On October 21, 2026, websites around the world started showing “maintenance mode” pages. Government, school, and hospital sites in Europe went offline on purpose, and companies like Nintendo took some services down too.</p><p>The reason was not a hack of “the internet”. It was two critical bugs in PHP, the software that runs behind a huge part of the web. Together they are called <strong>PHP2Shell</strong>, and they had been sitting in PHP’s code for more than ten years. AI helped find them.</p><h3>First, what is PHP?</h3><p>PHP is a programming language that runs on web servers. When you open a website, PHP is often the part that takes your request, talks to the database, and builds the page you see. WordPress, Drupal, and Joomla all run on it. Wiz Research says PHP powers about 70% of websites.</p><p>Think of a website as a hotel:</p><ul><li><strong>You</strong> are the guest walking in.</li><li><strong>The hotel’s rooms</strong> are the website’s data: accounts, orders, passwords.</li><li><strong>PHP is the receptionist.</strong> Every guest goes through the front desk first. The receptionist reads your request, checks the register, and sends you where you need to go.</li></ul><p>The website owner hires the receptionist, but does not train them. PHP comes ready-made. So if the receptionist has a bad habit, every hotel that uses them has the same problem. That is why this news is so big: the bug is in PHP itself, not in one website’s code.</p><h3>The receptionist’s two bad habits</h3><p><strong>Bad habit 1: Panicking when the desk is full</strong></p><p>When PHP runs out of memory, it stops what it is doing immediately. Sometimes it stops while a guest’s form is only half filled in. Later, during clean-up, PHP treats that half-filled form as if it were complete and correct. An attacker can send a specially built upload request that pushes PHP into exactly this situation. On servers using Apache with mod_php, this can happen before the website’s own code even runs.</p><p><strong>Bad habit 2: Not taking the old key back</strong></p><p>PHP keeps some notes from one guest’s visit after that guest has left and their space has been cleared. If the next guest is served by the same worker, PHP can use those stale notes before resetting them. It is like a hotel where the last guest’s room key still opens the door for the next person.</p><p><strong>Why it matters</strong></p><p>By using these habits, an attacker can make the server run their own commands. No username or password is needed. Wiz also says attackers could read passwords, tokens, and API keys from PHP’s memory.</p><p>The PHP security team rated both bugs <strong>9.2 out of 10</strong> (CVSS v4.0, “critical”). A score of 10.0 went around on social media before the release; the PHP Foundation says that number was incorrect.</p><h3>Hidden for over a decade, found with AI</h3><p>Every PHP version from <strong>5.4 up to 8.5.11</strong> is affected. Wiz says every PHP version released in the past decade carries at least one of the two bugs, and they survived years of manual security reviews.</p><p>So why now? Two teams found the problem using AI:</p><ul><li><strong>Calif:</strong> researcher Ryan Austin used OpenAI’s GPT-5.6 Sol during an August 2026 audit of PHP’s source code. The audit was funded by OpenAI under its “Patch the Planet” programme, which pairs AI models with human security experts to find and fix bugs in open-source software.</li><li><strong>Wiz Research</strong> independently found and reported both bugs, also with help from AI models. Wiz says proving the bugs could actually be exploited needed a chain of steps so complex it sat at the edge of what human researchers could practically do.</li></ul><p>Volker Dusch, a security engineer at the nonprofit that oversees PHP, called it the worst security problem the project has faced in more than ten years and compared it to Heartbleed, according to The Wall Street Journal.</p><p>The important point: AI did not create these bugs. They were already there. AI helped security researchers find them, so PHP could build a fix.</p><h3>What happened this week</h3><p>The plan was a quiet, coordinated fix. Instead, the bug IDs leaked a day early and attackers started probing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/561/1*yxhJ4jLcsyu_6yFVtqpHCA.png" /></figure><p>The EU, the US, Germany, Japan, Australia, the UK and the Netherlands all issued local warnings, according to <a href="https://therecord.media/php-script-vulnerabilities-warnings">The Record</a>.</p><h3>What you should do</h3><p><strong>If you run a website on PHP</strong></p><ul><li>Update PHP to <strong>8.5.12, 8.4.27, 8.3.36 or 8.2.35</strong>. If you install PHP through your operating system’s package manager, apply your vendor’s updated packages (<a href="https://it.uw.edu/help/servicestatus/php-cves/">UW-IT</a>).</li><li>On older branches: PHP released minimal source backports for 7.4, 8.0, and 8.1 (<a href="https://www.helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619/">Help Net Security</a>). watchTowr notes there are no patches for PHP 5 or 7 overall (<a href="https://watchtowr.com/intelligence/php-cve-2026-109618-faq/">watchTowr</a>).</li><li>On shared hosting, ask your provider to confirm the fix is <strong>installed</strong>, not just scheduled (<a href="https://www.malwarebytes.com/blog/bugs/2026/10/php2shell-why-websites-are-going-offline-for-emergency-security-updates">Malwarebytes</a>).</li><li>If you take a site down while patching, Google’s John Mueller advises returning HTTP 503 rather than letting pages time out or return 404/500 (<a href="https://www.seroundtable.com/google-php2shell-42353.html">Search Engine Roundtable</a>).</li></ul><p><strong>If your site is behind Cloudflare</strong></p><p>Cloudflare’s CTO Dane Knecht said PHP sites behind Cloudflare already have an initial set of firewall (WAF) rules for the known exploit paths, on every plan including Free (<a href="https://www.helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619/">Help Net Security</a>). Keep those rules on, but they are an extra layer only. Because the bug lives in PHP itself, you still have to update PHP.</p><p><strong>If you are just browsing</strong></p><p>The fix is applied on web servers, not on your phone or laptop. Some sites may stay in maintenance mode for a while as their teams patch (<a href="https://www.malwarebytes.com/blog/bugs/2026/10/php2shell-why-websites-are-going-offline-for-emergency-security-updates">Malwarebytes</a>).</p><h3>The lesson for testers: test the failure path</h3><p>Look at where both bugs lived. Not in login, search, or checkout. They lived in what PHP does <strong>when things go wrong</strong>: running out of memory, hitting a fatal error, cleaning up between requests.</p><p>Most test suites are built around the happy path. We check that the right input gives the right output. Far fewer tests ask:</p><ul><li>What happens when the server runs out of memory halfway through a request?</li><li>What state is left behind after a crash, and does the next request see it?</li><li>What does the system do with a huge or oddly formatted upload?</li></ul><p>That is exactly the kind of corner where this bug hid for over a decade. Wiz says AI models lowered the barrier to proving bugs like this are exploitable, so these corners deserve more of our testing time, not less.</p><p>One small step this week: add one “what if it breaks halfway?” test to your suite.</p><h3>Myths vs facts</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/440/1*3dH_VRbtetoCbDFQAQl9gA.png" /></figure><p>If you want to stay updated on tech &amp; AI updates, follow me on other platforms:</p><p><strong>Anshita Bhasin</strong><br><a href="https://www.instagram.com/abautomationhub_/">Instagram</a> | <a href="https://www.linkedin.com/in/anshita-bhasin/">LinkedIn</a> | <a href="https://twitter.com/anshita__bhasin">Twitter</a> | <a href="https://www.youtube.com/channel/UCtNHEcO_c9qpaUqIYSb98Rg">Youtube</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a6e13716dba1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Anthropic Stopped Using Git for Its Monorepo. Here’s Why.]]></title>
            <link>https://medium.com/@anshita.bhasin/anthropic-stopped-using-git-for-its-monorepo-heres-why-d958c2ea1b17?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/d958c2ea1b17</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[version-control]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 18:31:30 GMT</pubDate>
            <atom:updated>2026-10-15T18:31:30.405Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VcNTkysr1x4vzdGeuXS-zg.png" /></figure><p><em>What happens when AI writes code faster than your tools can keep up?</em></p><p>Most of us have never thought of Git as a bottleneck. It’s just there. You pull, you commit, you push, and you move on with your day.</p><p>So when Anthropic published an engineering post on October 13 explaining why they had moved their entire monorepo off Git, I had to read it twice.</p><p>The short version is this. In May 2026, Anthropic was merging around 2,500 code changes a day. By September, that number was over 30,000. A day. Git, along with all the workarounds they had built around it, simply couldn’t keep up. So a small team built a new version control system called Scion, and moved everything over to it in five months.</p><p>I want to walk through why this happened, because I think the reason matters more than the tool itself.</p><h3>When code stops being the slow part</h3><p>For as long as most of us have been in this industry, writing code has been the expensive part. Everything else : the builds, the pipelines, the reviews, was designed around the pace at which humans write and ship it.</p><p>Anthropic’s post opens with a line that flips that around: “At Anthropic, coding is no longer the bottleneck to building systems or doing research.”</p><p>Claude now writes so much of their code, and so quickly, that the slow part has moved somewhere else. In their words, the new bottleneck is “all of the developer infrastructure built for human scale: source control, builds, and CI.”</p><p>And they don’t see it calming down. They describe the demand for merging code as “practically infinite”, and say that every bottleneck they remove just leads to even more code being written.</p><h3>Git was never built for this</h3><p>To be fair to Git, Anthropic is careful not to blame it. Git is decentralized by design. Every clone is a complete copy of the repository, so you can check history, run a diff, or switch branches while sitting on a plane with no Wi-Fi. For almost every team out there, that is exactly what you want, and it’s a big reason Git became the standard.</p><p>The trouble is scale. Anthropic’s main branch history alone is almost 100 GB. A single compressed working copy is over 25 GB. Their CI checks out code roughly 25 million times a day, and that’s before you count developer environments, agent sandboxes, and research jobs. If every one of those checkouts pulled the repo the normal way, it would add up to more than 600 PB of Git traffic every day.</p><p>They didn’t jump straight to building something new. They tried the usual tricks first: shallow clones, sparse checkouts, saved snapshots of the repo, and internal Git mirrors. At one point they were running several hundred of those mirrors just to handle the load. It bought them time, but as the merge rate kept climbing, keeping all of it running got harder and harder.</p><h3>The question that changed everything</h3><p>At some point the team stepped back and noticed something about how they were really using Git.</p><p>Like most companies, they treated it as a central system with one source of truth. And nearly all of the “users” pulling from that repo weren’t people on laptops. They were AI agents and automated jobs, which are always connected. The very feature that makes Git great, working fully offline with a full copy, was the thing working against them.</p><p>So they asked a simple question:</p><blockquote><em>If we designed source control with the fact that agents are always online in mind from the start, what would it look like?</em></blockquote><p>The answer to that question is <strong>Scion</strong>.</p><h3>So what is Scion?</h3><p>Scion is Anthropic’s in-house version control system. They didn’t design it from a blank page. They borrowed lessons from the systems Google and Meta built for their own huge repos, <a href="https://en.wikipedia.org/wiki/Piper_(source_control_system)">Piper</a> and <a href="https://sapling-scm.com/">Sapling</a>.</p><p>The biggest change is where the work happens. With Git, your machine does almost everything: it stores the history, works out the diffs, creates the commits. With Scion, the server does that heavy lifting. Commits, merges, rebases, history, and blame all happen on the server. The client keeps a small cache and only downloads the files it actually needs, when it needs them.</p><p>That has a nice side effect. Because the server can answer questions directly, every query is also available through an API. A CI job that wants to know what changed in a commit can just ask, without checking out the repo at all. And since commits are created on the server, there’s nothing to push.</p><p>The other big idea is that nothing lives in one place. No single machine holds a full copy of the repo, not even the server. The data is spread out, and every part of the system can scale on its own.</p><h3>Did it pay off?</h3><p>The numbers say yes.</p><p>Anthropic reports that the infrastructure behind Scion is 200 times smaller in CPU and memory than the Git mirror setup it replaced. CI and research jobs now check out code in seconds rather than minutes, about two orders of magnitude faster. CI alone got back 25% of its CPU capacity. Every layer of caching hits more than 99% of the time, and there’s no repository maintenance left for anyone to babysit.</p><p>When you’re doing tens of millions of checkouts a day, that kind of speed-up frees up what Anthropic calls “a sizable amount of capacity across the fleet.”</p><h3>Built with Claude, steered by people</h3><p>There’s a nice irony in this story. Claude’s ability to write code is what created the problem, and it’s also a big part of how the team solved it so quickly.</p><p>The example that stuck with me was the pipeline that moves data from Git into Scion. The team had written the first version mostly by hand and were happy with it. Then they gave it to an early internal snapshot of <a href="https://www.anthropic.com/claude/mythos">Claude Mythos Preview</a> and asked it to rewrite the whole thing, with no further instructions. The new version used 20 times less memory and ran four times faster, while keeping all the rules the original depended on.</p><p>The team itself was tiny. Seven people built Scion, a new code review tool called Antique, and about half a dozen related systems, then migrated the whole monorepo, all in five months. Anthropic says a project like this would once have taken a company years and many more engineers.</p><p>What I liked is that the post doesn’t pretend the engineers stopped mattering. Everyone on the team now rates Claude as better than them at writing code. But they were the ones making the design calls, and very often their job was to tell Claude “no”, for example, when it tried to over-engineer something because of a problem that didn’t really exist. They still review every change. They just focus on what the change is trying to do instead of reading every line.</p><h3>What stood out to me as a QA person</h3><p>Honestly, the most interesting part of the post for me wasn’t about Git at all. It was about testing.</p><p>Anthropic writes that a traditional merge queue “doesn’t make much sense at agent scale”, and then asks a question I keep coming back to: “What does CI even mean when hundreds or even thousands of changes can merge while tests are running?”</p><p>At their volume, they couldn’t make the industry-standard approach work, which is locking the whole repo on the merge queue while tests run. So they shifted their effort to post-merge and pre-deploy automated testing. Yes, that means changes can sometimes conflict. But Claude can usually fix those conflicts faster than a human could even spot them.</p><p>The other thing I noticed is that faster code didn’t mean less testing. It meant more. Because Claude could write so much code so quickly, the team built simulations of every tricky infrastructure and source control situation they had ever run into, along with fuzzing and rate limiting. Anthropic calls Scion one of the best-tested codebases they’ve worked on. For anyone who has ever had to cut testing because of a deadline, that’s a very different picture.</p><h3>Should you switch?</h3><p>No, and you couldn’t if you wanted to. Scion is an internal system. Anthropic hasn’t released it, and they’re very direct about why:</p><blockquote><em>Scion is by no means a suitable replacement for git for general use.</em></blockquote><p>It only makes sense for truly enormous monorepos, the kind with tens of thousands of engineers or swarms of agents writing code. It works so differently from Git that a company would have to change every part of its development lifecycle to adopt it, including code review, code search and CI. And Anthropic’s versions of those pieces are built specifically for them.</p><p>One small thing worth clearing up, because Git and GitHub often get mixed up: Scion is an alternative to Git, the version control tool. It’s not a competitor to GitHub or GitLab, which are platforms that host Git repositories.</p><p>What Anthropic does say is that this kind of problem is coming for others. They expect it to become more common over the next six to twelve months, as more teams use AI coding tools.</p><p>Rather than release Scion, they shared the thinking behind it so Git platforms can borrow the parts that work, like answering more questions on the server side. They also say they want to work with GitHub, GitLab, and the open source community on it.</p><h3>Conclusion :</h3><p>Scion isn’t going to replace Git, and Anthropic is the first to say so. But I don’t think that’s the real takeaway.</p><p>The real story is what happens when AI starts writing code at a pace no human team ever could. The tools we’ve relied on for years, the ones built around human speed, start to show their limits.</p><p>Anthropic calls source control “just one example of this”, and says the project made them “take a very hard look at other workflows designed for human scale.”</p><p>Most of us aren’t merging 30,000 changes a day. But if you work in QA, DevOps, or platform engineering, it’s worth asking which parts of your pipeline quietly assume a human is on the other end.</p><p><em>I also made a short video telling this story. You’ll find it on ABAutomationHub on YouTube and Instagram.</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FX8itg0H-aOo&amp;type=text%2Fhtml&amp;schema=youtube&amp;display_name=YouTube&amp;src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FX8itg0H-aOo%3Ffeature%3Doembed" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/61f2382a62bcbcf843a91c967911a6e8/href">https://medium.com/media/61f2382a62bcbcf843a91c967911a6e8/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d958c2ea1b17" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Jev AI, Seen Through a Tester’s Eyes]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@anshita.bhasin/jev-ai-seen-through-a-testers-eyes-cd70cfda241d?source=rss-cf4a02363eb1------2"><img src="https://cdn-images-1.medium.com/max/1600/1*ngDds7-MCWC6faMu85cldQ.png" width="1600"></a></p><p class="medium-feed-snippet">A model that doesn&#x2019;t chat; it decides. What it is, what it claims, and how QA teams can use it.</p><p class="medium-feed-link"><a href="https://medium.com/@anshita.bhasin/jev-ai-seen-through-a-testers-eyes-cd70cfda241d?source=rss-cf4a02363eb1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@anshita.bhasin/jev-ai-seen-through-a-testers-eyes-cd70cfda241d?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd70cfda241d</guid>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[jevai]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 19:45:26 GMT</pubDate>
            <atom:updated>2026-10-09T19:45:26.883Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[CI/CD for QA: Understanding What Really Happens to Testing When Software Delivery Becomes Automated]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@anshita.bhasin/ci-cd-for-qa-understanding-what-really-happens-to-testing-when-software-delivery-becomes-automated-cb9584d0a6c5?source=rss-cf4a02363eb1------2"><img src="https://cdn-images-1.medium.com/max/957/1*ZiffudISelLets4jSKeNkg.png" width="957"></a></p><p class="medium-feed-snippet">For a long time, software testing was treated as a stage that happened after development.</p><p class="medium-feed-link"><a href="https://medium.com/@anshita.bhasin/ci-cd-for-qa-understanding-what-really-happens-to-testing-when-software-delivery-becomes-automated-cb9584d0a6c5?source=rss-cf4a02363eb1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@anshita.bhasin/ci-cd-for-qa-understanding-what-really-happens-to-testing-when-software-delivery-becomes-automated-cb9584d0a6c5?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb9584d0a6c5</guid>
            <category><![CDATA[cicd]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[qa]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 04:10:45 GMT</pubDate>
            <atom:updated>2026-09-22T04:10:45.266Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Write Prompts the Way You Actually Think, Introducing “ShapeMyPrompt”]]></title>
            <link>https://medium.com/@anshita.bhasin/write-prompts-the-way-you-actually-think-introducing-shapemyprompt-84b89ca00bed?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/84b89ca00bed</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[chrome-extension]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Wed, 19 Aug 2026 17:39:23 GMT</pubDate>
            <atom:updated>2026-08-19T17:39:23.633Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xMeul2Z0p9bkKBJFCVX2pQ.png" /></figure><p>You have a good idea in your head. Then you sit down to type it into ChatGPT, Claude, Gemini, or Perplexity, and half your time goes into finding the right English words instead of the idea itself. The thought breaks. The prompt that comes out is never as sharp as what you had in mind.</p><p>That is the exact problem <strong>ShapeMyPrompt</strong> solves.</p><h3>The Idea</h3><p>Type in whatever language feels natural to you: Hinglish, Tamil-English, Spanglish, or broken English. ShapeMyPrompt converts it into a clear, structured English prompt, and that refined prompt is ready to use instantly. No copying, no pasting, no switching tabs.</p><h3>What It Does</h3><ul><li>Converts your thoughts into clear English</li><li>Turns them into structured, effective prompts</li><li>Makes the improved prompt usable instantly, without copying, pasting, or switching tabs</li></ul><h3>How It Works</h3><ol><li>Add the ShapeMyPrompt Chrome extension</li><li>Open ChatGPT, Claude, Gemini, or Perplexity</li><li>Type your request naturally, in whatever language feels natural to you</li><li>Click the ShapeMyPrompt button inside the prompt box</li></ol><p>That’s it. Your refined prompt is ready to send.</p><p>Using mobile or don’t have Chrome? A web version is also available — paste your text, shape the prompt, and copy the result.</p><h3>Why It’s Free</h3><p>I built ShapeMyPrompt as a contribution from #<a href="https://www.youtube.com/@ABAutomationHub/videos"><strong>ABAutomationHub</strong></a> to the community that has supported my journey. This is an early version, and your feedback will decide what gets built next.</p><h3>Try It</h3><ul><li>Chrome Extension: <a href="https://chromewebstore.google.com/detail/opcpkdenckimhohekpbemidlndchljae?utm_source=item-share-cb">Link</a></li><li>Web Version: <a href="https://shapemyprompt.com/">Link</a></li></ul><p>Try it. Share it. Tell me what’s missing.</p><p><em>#ShapeMyPrompt #PromptEngineering #GenerativeAI #ChromeExtension #ArtificialIntelligence #ABAutomationHub</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=84b89ca00bed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI for QA Engineers: New Live Batch Starts 25 July]]></title>
            <link>https://medium.com/@anshita.bhasin/ai-for-qa-engineers-new-live-batch-starts-25-july-a65ddf9a735f?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a65ddf9a735f</guid>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Fri, 17 Jul 2026 08:56:01 GMT</pubDate>
            <atom:updated>2026-07-17T08:56:01.850Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A hands-on live course for working QA engineers who want to actually use AI in their testing work, not just talk about it.</em></p><p>Earlier this year, I ran the first batch of my live course, AI for QA. Working QA engineers from different companies and countries joined, sat through live sessions, built things hands-on, and presented their own projects on Demo Day.</p><p>The most common feedback I got was some version of this: “I had read about AI everywhere, but this was the first time I actually used it in my own testing workflow.”</p><p>That is exactly the gap this course exists for. And now the next batch is starting on <strong>25 July</strong>.</p><p>This post covers everything: what the course teaches, how it runs, who it is for, and who it is not for.</p><h3>Why this course exists</h3><p>If you work in QA today, you have probably noticed two things happening at the same time.</p><p>First, every job description now mentions AI in some form. Teams expect QA engineers to know how to use AI tools, not just have an opinion on them.</p><p>Second, most of the content available online is either too shallow (generic “10 ChatGPT prompts for testers” lists) or too far from real QA work (ML theory courses built for data scientists).</p><p>There is very little in the middle: practical, hands-on training built specifically for QA engineers, using real testing scenarios, taught live so you can ask questions when you get stuck.</p><p>That middle is what this course is.</p><h3>What you will learn</h3><p>This is a live, hands-on course. Every topic is taught by doing, on screen, with real examples from QA work. The curriculum covers:</p><p><strong>Foundations that actually matter</strong></p><ul><li>Tokens, context windows, and why your prompts behave the way they do</li><li>Context engineering: how to structure what you give an AI so you get reliable output instead of random output</li></ul><p><strong>Working with Claude for QA tasks</strong></p><ul><li>Using Claude for day-to-day QA work: test case generation, test data, bug analysis, documentation</li><li>Claude Projects and Skills, and how to set them up for your own testing context</li><li>Claude Code: using it hands-on for automation work</li></ul><p><strong>MCP (Model Context Protocol)</strong></p><ul><li>What MCP is and why it matters for testers</li><li>Using the official Playwright MCP server to drive browser automation through AI</li><li>Building your own custom MCP server, step by step</li></ul><p><strong>Testing the AI itself</strong></p><ul><li>Promptfoo: evaluating and testing prompts systematically, the QA way</li></ul><p><strong>Beyond one tool</strong></p><ul><li>GitHub Copilot in a QA workflow</li><li>Running local LLMs with Ollama, so you understand what runs on your machine versus in the cloud</li></ul><p><strong>Build your own practice environment</strong></p><ul><li>Building your own website from scratch to use as your personal practice ground for testing</li><li>Taking that website live, so you have a real deployed application to test against, not just localhost</li><li>Creating your own database and connecting it to your application</li></ul><p>This part matters more than it sounds. Most testers practice on the same public demo sites everyone else uses. In this course, you build and deploy your own application with its own database. That gives you a real system you fully control, to practice UI testing, API testing, and AI-assisted workflows on.</p><p><strong>Demo Day</strong></p><ul><li>The course ends with a Demo Day where you present what you built. Not a certificate for attendance. An actual project you made yourself.</li></ul><p>Everything taught in this course is based on official documentation and hands-on verified workflows. No secondhand claims, no “AI can probably do this” content.</p><h3>How the batch runs</h3><ul><li><strong>Start date:</strong> 25 July</li><li><strong>Schedule:</strong> Saturday and Sunday, weekend-only</li><li><strong>Time:</strong> 8:00 to 9.30:00 AM IST (1.5-2 hours per session)</li><li><strong>Format:</strong> Live sessions on screen share, fully hands-on</li><li><strong>Recordings:</strong> Every session is recorded, so if you miss a live class, you do not fall behind</li></ul><p>The weekend-morning format is deliberate. Most people joining are working professionals. Two consecutive weekend sessions keep the momentum going without eating into your work week.</p><h3>Who this course is for</h3><ul><li>QA engineers (manual or automation) who want to add AI skills to their real work, not just their resume</li><li>Testers who have used ChatGPT casually but never gone beyond basic prompting</li><li>Automation engineers curious about Claude Code, MCP, and AI-assisted test development</li><li>Anyone in QA who keeps seeing “AI” in job descriptions and wants hands-on ability, not just awareness</li></ul><p>You do not need any AI or ML background. You do need to be willing to do the hands-on work between sessions.</p><h3>Who this course is NOT for</h3><p>I would rather be honest here than have someone join and feel it was not for them.</p><ul><li>If you want deep machine learning theory or model training, this is not the course</li><li>If you want a recorded, watch-at-your-own-pace course, this is live and structured around live participation</li><li>If you expect AI to replace your testing skills, this course teaches the opposite: AI as a tool in the hands of a skilled tester</li></ul><h3>Common questions</h3><p><strong>Do I need to know coding?</strong> Basic familiarity helps for the automation-focused sessions, but the course is built for QA engineers, including manual testers. Every hands-on exercise is walked through live, step by step.</p><p><strong>What if I miss a session?</strong> Recordings are provided for every session.</p><p><strong>What tools do I need?</strong> Everything you need is covered in the setup session at the start. You will not be asked to buy a stack of paid tools.</p><p><strong>Is this the same as Batch 01?</strong> The core curriculum is the same practitioner-focused approach, updated wherever the tools themselves have changed since the last batch.</p><h3>How to join</h3><p>The new batch starts <strong>Saturday, 25 July</strong>.</p><p>Registration link: <a href="https://www.youtube.com/redirect?event=comments&amp;redir_token=QUFFLUhqa2dEaUEwVHNvczlLeDczajRYNklmN0R5UklnUXxBQ3Jtc0tsdE9vRW80RzU5RkpHaThQZUQ4Mm85UV90NDhnQWlvdWVzN25LQnE2U2dzS1NWRDYwQ0VQNG01R3AtV2k5VlJaZ0ZZN01BWkdTQXJuMlJyaGRKTG5CV1JsQks2aFpNNDVINE16Q1o5Y2JhZ1F1MkhqNA&amp;q=https%3A%2F%2Fforms.gle%2FTfAjJQ1EU3xtxu3x9">https://forms.gle/TfAjJQ1EU3xtxu3x9</a></p><p>If you have questions before registering, drop them in the comments or reach out to me directly.</p><p>See you in class.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FXuHYwIaXGQ0%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DXuHYwIaXGQ0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FXuHYwIaXGQ0%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9353700fdaa4dc91528279bd102fe6a7/href">https://medium.com/media/9353700fdaa4dc91528279bd102fe6a7/href</a></iframe><p><em>I run ABAutomationHub, where I create content on QA, automation, and AI for testers. This course is the live, hands-on version of everything I have been building and teaching in that space.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a65ddf9a735f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM-as-a-Judge: How AI Evaluates AI]]></title>
            <link>https://medium.com/@anshita.bhasin/llm-as-a-judge-how-ai-evaluates-ai-1d0411a3c16d?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d0411a3c16d</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Sat, 04 Jul 2026 09:25:51 GMT</pubDate>
            <atom:updated>2026-07-04T09:25:51.330Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*95vPHSO1VBtjE1V2XSNp0w.png" /></figure><h3>The Problem With Evaluating AI at Scale</h3><p>You built an AI feature. It works, most of the time. But how do you actually know it’s good? Manual human review is accurate, but doesn’t scale. If you have 100,000 prompts and 5 model versions to compare, that’s <strong>500,000 responses</strong> to evaluate. No team can do that sustainably.</p><p>That’s exactly why <strong>LLM-as-a-Judge</strong> exists.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/929/1*FGPnwH59NXOFVff-DmJ_zQ.png" /></figure><h3>What is LLM-as-a-Judge?</h3><p>LLM-as-a-Judge is an evaluation method where <strong>one language model evaluates the output of another</strong>, using a defined scoring rubric, instead of asking a human to review every response.</p><p><strong>Simple example:</strong></p><blockquote>User Prompt: “Explain photosynthesis to a 10-year-old.”</blockquote><p>Your model generates a response. Instead of sending it to a human, you send it to a second LLM with a prompt like:</p><blockquote>Evaluate this response on:<br> — Accuracy (1–5)<br> — Clarity (1–5)<br> — Age appropriateness (1–5)</blockquote><blockquote>Give a score for each and explain your reasoning.</blockquote><p>That second model is the <strong>judge</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/771/1*y4FM_5JqhL79-whYEGOstg.png" /></figure><h3>The Research Behind It</h3><p>This isn’t a new or experimental idea. The foundational paper, <strong>“Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,”</strong> was published at <strong>NeurIPS 2023</strong> by researchers at LMSYS (UC Berkeley).</p><blockquote><em>Strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement seen between two humans.</em></blockquote><blockquote><em>— </em><a href="https://arxiv.org/abs/2306.05685"><em>arXiv:2306.05685, NeurIPS 2023</em></a></blockquote><p>They introduced two benchmarks to validate this at scale:</p><ul><li><strong>MT-Bench</strong> — a multi-turn question set for testing chat assistants</li><li><strong>Chatbot Arena</strong> — a crowdsourced platform where humans vote on model quality</li></ul><h3>Three Ways to Use an LLM Judge</h3><p>OpenAI’s official evaluation documentation outlines three grading approaches:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/809/1*yANMzESUJ0cmrYSf2BCavA.png" /><figcaption><a href="https://developers.openai.com/api/docs/guides/evaluation-best-practices">Open A Eval Best Practices</a></figcaption></figure><h3>Real-World Usage: What Anthropic Actually Does</h3><p>Anthropic published the details of their LLM-as-a-Judge implementation in their engineering blog.</p><h4><strong>Key finding from Anthropic:</strong></h4><blockquote><em>“A single LLM call with a structured prompt outputting scores from 0.0–1.0 and a pass/fail grade was the most consistent approach — and aligned best with human judgements.”</em></blockquote><blockquote><em>— </em><a href="https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents"><em>Anthropic Engineering Blog</em></a></blockquote><p><strong>Their specific recommendation:</strong> Grade each quality dimension with its own isolated LLM judge call, not all dimensions bundled into one prompt.</p><h4><strong>Real example from Anthropic’s agent evals:</strong></h4><blockquote>For a customer support agent handling refunds:<br>“Was the agent’s explanation grounded in the results of the fetch_policy tool?&quot;</blockquote><p>That single, narrow question gets its own judge call. Not mixed with five other criteria. This specificity is what makes scores reliable and actionable.</p><h3>Where It Works and Where It Doesn’t</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/1*1iJjeDrdS4t7vwHTB9kSMw.png" /></figure><p><strong>Rule of thumb:</strong> If the answer can be verified with code, use code. LLM judges are for subjective, open-ended quality assessment.</p><h3>Known Biases And How to Reduce Them</h3><p>LLM judges are not perfect. Research has documented four consistent biases:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/801/1*roFLfVj4cIVbbOlY1aengA.png" /><figcaption><a href="https://arxiv.org/html/2406.07791v7">Judging the Judges</a></figcaption></figure><p><strong>How to reduce bias:</strong></p><ul><li>Use pass/fail or pairwise formats — more reliable than numeric scales</li><li>Add chain-of-thought: ask the judge to reason <em>before</em> scoring</li><li>Use reference-guided grading when a gold standard exists</li><li>Periodically validate judge scores against human reviewers</li></ul><h3>What a Real Evaluation Pipeline Looks Like</h3><p>Most teams don’t use LLM-as-a-Judge in isolation. Each layer catches different failure types:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/789/1*Vyj6hPPJzm5FzWJF-AuX_g.png" /></figure><h3>What OpenAI and Anthropic Actually Say</h3><blockquote><strong><em>OpenAI</em></strong><em> (official eval documentation):<br>“Using models to judge output is cheaper to run and more scalable than human evaluation. No strategy is perfect. The quality of LLM-as-Judge varies depending on problem context.”</em></blockquote><blockquote><a href="https://platform.openai.com/docs/guides/evals/llm-as-a-judge"><em>OpenAI Evaluation Best Practices →</em></a></blockquote><blockquote><strong><em>Anthropic</em></strong><em> (engineering blog):<br>“LLM-as-judge graders should be closely calibrated with human experts to ensure minimal divergence between human and model grading.”</em></blockquote><blockquote><a href="https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents"><em>Demystifying Evals for AI Agents →</em></a></blockquote><p>Neither company relies on LLM judges exclusively. Both emphasize human calibration as the validation layer.</p><h3>The Simplest Way to Think About It</h3><p>Think of a school exam.</p><p>A student writes the answer. The teacher grades it, using a rubric, assigning marks, and explaining the score. In LLM-as-a-Judge, <strong>the teacher is another AI model</strong>. It reads the answer, follows a scoring rubric, gives a score, and explains its reasoning.</p><p>The rubric is what makes it reliable. Without a clear rubric, the judge is just guessing.</p><h3>Conclusion:</h3><p>If you’re shipping AI features without systematic evaluation, you’re guessing. LLM-as-a-Judge is one of the most practical tools available to fix that when used correctly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*cRk4trUw5GXwFHpfJJ8QGQ.png" /></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FNbsJlRpnpKw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FNbsJlRpnpKw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FNbsJlRpnpKw%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a333eb3b94076de3e14f5b6c1c4357b1/href">https://medium.com/media/a333eb3b94076de3e14f5b6c1c4357b1/href</a></iframe><p><em>Hope this helped you understand how LLM-as-a-Judge actually works. If you want hands-on training with AI in QA specifically, I run a focused batch called </em><strong><em>AI for QA</em></strong><em> where we go deep into practical techniques. If that sounds useful, the link is below. See you there.</em></p><p><strong>👉 </strong><a href="https://youtu.be/XuHYwIaXGQ0?si=qUCfKi6z9BVW-b6P"><strong>Join AI for QA Batch</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d0411a3c16d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deterministic vs Non-Deterministic AI]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@anshita.bhasin/deterministic-vs-non-deterministic-ai-7989fa2bdbaf?source=rss-cf4a02363eb1------2"><img src="https://cdn-images-1.medium.com/max/1470/1*njZf7Zy6EquMgMuStHaYBw.png" width="1470"></a></p><p class="medium-feed-link"><a href="https://medium.com/@anshita.bhasin/deterministic-vs-non-deterministic-ai-7989fa2bdbaf?source=rss-cf4a02363eb1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@anshita.bhasin/deterministic-vs-non-deterministic-ai-7989fa2bdbaf?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/7989fa2bdbaf</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[qa]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Fri, 26 Jun 2026 17:16:07 GMT</pubDate>
            <atom:updated>2026-06-27T08:03:33.620Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ADLC: Why SDLC Needs an Upgrade for AI Agents (and What’s Emerging Instead)]]></title>
            <link>https://medium.com/@anshita.bhasin/adlc-why-sdlc-needs-an-upgrade-for-ai-agents-and-whats-emerging-instead-65c126ec9029?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/65c126ec9029</guid>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Mon, 04 May 2026 05:02:12 GMT</pubDate>
            <atom:updated>2026-05-04T05:02:12.479Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qgJ-jWV-NbrPjp_vxFdRlg.png" /></figure><p>We’re trying to build non-deterministic systems using a lifecycle designed for deterministic software. That gap is starting to break things</p><p>SDLC works well for deterministic software. You define behavior upfront, implement it, test it, deploy it, and expect consistent outcomes. <strong><em>AI agents don’t behave like that.</em></strong></p><p>They introduce non-determinism, context sensitivity, and evolving behavior, forcing teams to rethink how systems are designed, tested, and governed.. That’s where the idea of an <strong>Agentic Development Lifecycle (ADLC)</strong> is starting to emerge.</p><p><strong><em>Not as a replacement for SDLC, but as an extension of it.</em></strong></p><h3>Why Traditional SDLC Struggles with Agents</h3><p>SDLC assumes predictable execution paths:</p><blockquote><strong>If X happens → system produces Y</strong></blockquote><p>That assumption breaks with AI systems. Modern AI systems are probabilistic, context-sensitive, and continuously evolving.</p><p>Even foundational AI lifecycle research highlights that these systems require <strong>ongoing monitoring, retraining, and adaptation</strong>, not a one-time “build and ship” mindset.</p><p>Agentic systems take this further.</p><p>They:</p><ul><li>Call APIs</li><li>Interact with external tools</li><li>Make multi-step decisions</li></ul><p>And failures aren’t just bugs. They can become:</p><ul><li>Incorrect decisions</li><li>Security risks</li><li>Financial or Operational impact</li></ul><h3>What ADLC Actually Means (Today)</h3><p>The Agentic Development Lifecycle is a framework that inverts the assumptions of traditional software development.</p><p>Instead of: <strong><em>specification first → code → test → deploy locked</em></strong></p><p>You get: <strong><em>intent first → design constraints → generate with validation in parallel → observe and correct continuously</em></strong></p><p>Let me break the core phases (this matters for how QA and testing work):</p><p><strong>Phase 0: Intent &amp; Problem Discovery, not</strong> requirements gathering. You’re discovering what problem the agent actually needs to solve, stated as an observable outcome, not a technology goal. What does success look like? What are the hard constraints? Where can the agent make mistakes without catastrophic failure?</p><p><strong>Phase 1: Agent Design.</strong> Define the high-level objective, the guardrails that govern autonomy, which workflow steps need non-deterministic reasoning, and which must stay deterministic (financial calculations stay hardcoded, not left to LLM judgment).</p><p><strong>Phase 2: Generation &amp; Validation in Parallel.</strong> This is the critical difference from SDLC. The agent generates code or decisions. Simultaneously, validation runs. The agent itself identifies edge cases, flags regressions, and surfaces failure modes before a human reviews anything.</p><p>The measure of a healthy Validate phase isn’t test coverage percentage. It’s bet resolution rate: did this output resolve the problem or create new ones?</p><p><strong>Phase 3: Govern.</strong> The phase ADLC adds that SDLC doesn’t have. Human judgment layer. Not “is this technically correct” — but “does this output serve the business intent? Does the signal support advancing this?”</p><p><strong>Phase 4: Observe &amp; Continuous Correction</strong> You’re not just watching uptime anymore. You’re watching reasoning traces. The agent’s thought process. Which tools it called? In what order? Why does it choose one path over another?</p><p>If an agent starts reading secrets and then making outbound API calls, you catch it. If behavior drifts from what you validated in testing, you know immediately</p><h3>What does this change for QA and Testing</h3><p>If you’re a QA engineer or test automation engineer, here’s what changes for you:</p><p>1. You’re not writing test cases the way you used to. You’re building evaluation harnesses. Continuous test suites that grow as the system encounters new production failure cases.</p><p>2. Validation is not a gate at the end of the pipeline. It runs alongside generation from day one.</p><p>3. You’re measuring system behavior across multi-step workflows where errors compound. A failure in step 2 might not surface until step 5. Your test strategy needs to catch that.</p><p>4. You’re not testing for exact output matching. You’re testing for behavioral quality and alignment with intent. Did the agent stay within its guardrails? Did it reason correctly across the workflow?</p><p>5. Security testing shifts. You’re not just scanning code. You’re monitoring which tools the agent accesses, which MCP servers it connects to, and whether its tool-call patterns look anomalous</p><h3>Real-World Deployments, Failures, and Lessons</h3><p>This shift isn’t theoretical. We already have real examples that show where traditional thinking breaks.</p><h4>Microsoft Tay: No Guardrails, Instant Failure</h4><p><strong>What happened:</strong><br> Tay learned from users on Twitter and began producing offensive content within 24 hours.</p><p><strong>Lesson: </strong><em>Systems that learn from live environments without constraints will drift — fast.</em></p><h4>Microsoft Bing AI: Behavior Degrades Over Time</h4><p><strong>What happened:</strong><br> Early versions produced erratic and sometimes hostile responses in long conversations.</p><p><strong>Lesson: </strong><em>Multi-step interactions introduce failure modes not visible in short tests.</em></p><h4>Tencent Chatbots: Alignment Failures</h4><p><strong>What happened:</strong><br> Chatbots produced politically sensitive responses and were shut down.</p><p><strong>Lesson: </strong><em>Alignment is not one-time; it must be continuously enforced.</em></p><h3>What These Examples Prove</h3><p>Across all cases:</p><ol><li>Failures happen in production, not just in testing.</li><li>Most failures are system design issues, not model limitations.</li><li>Behavior breaks over time. What works in demo fails at scale.</li><li>When agents have autonomy, the stakes are higher.</li></ol><h3>Important Clarification</h3><p>Adding Copilot to your workflow isn’t ADLC.</p><p>Bolting an AI test tool onto QA isn’t ADLC.</p><p>Those are productivity improvements inside an unchanged operating model. Valuable, not transformative.</p><p>ADLC requires a different structure for how work flows when agents are the primary execution unit. If your organization can’t commit to parallel generation and validation, continuous behavioral monitoring, and shifted human judgment to governance, you’re operating SDLC with acceleration. Not ADLC.</p><h3>What Matters Right Now</h3><p>If you’re building with Claude, Claude Code, or any agentic system, understanding ADLC shapes how you:</p><ul><li>Define success metrics</li><li>Write evaluation prompts</li><li>Build monitoring</li><li>Structure your testing approach</li><li>Hand off work to team members</li></ul><p>You don’t need to adopt the full framework tomorrow. But knowing the shape of how agentic systems should be built — as non-deterministic, continuously learning systems- changes what you build today.</p><p>The software engineering industry spent decades figuring out how to build deterministic systems reliably. We’re at the beginning of figuring out how to do it for agentic ones.</p><p>ADLC is the first structured answer to that question.</p><h3>Conclusion:</h3><p>The industry spent decades mastering deterministic systems.</p><p>Now we’re building systems that generate, adapt, and behave probabilistically.</p><p>ADLC isn’t a finalized standard yet. It’s an emerging direction.</p><p><strong>Different failures. Same root problem.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=65c126ec9029" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meta’s AI Got It Wrong. The Fix? Classic QA. ( Case Study)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@anshita.bhasin/metas-ai-got-it-wrong-the-fix-classic-qa-case-study-f4f8438e3673?source=rss-cf4a02363eb1------2"><img src="https://cdn-images-1.medium.com/max/977/1*dnEyuGhXWGM_nALiCMtICQ.png" width="977"></a></p><p class="medium-feed-snippet">AI is supposed to reduce the need for QA.</p><p class="medium-feed-link"><a href="https://medium.com/@anshita.bhasin/metas-ai-got-it-wrong-the-fix-classic-qa-case-study-f4f8438e3673?source=rss-cf4a02363eb1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@anshita.bhasin/metas-ai-got-it-wrong-the-fix-classic-qa-case-study-f4f8438e3673?source=rss-cf4a02363eb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4f8438e3673</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[software-testing]]></category>
            <dc:creator><![CDATA[Anshita Bhasin]]></dc:creator>
            <pubDate>Wed, 29 Apr 2026 06:09:44 GMT</pubDate>
            <atom:updated>2026-04-29T06:12:27.167Z</atom:updated>
        </item>
    </channel>
</rss>