<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>Blog · Metanaps — Anthony Carbonnier</title>
    <link>https://metanaps.com/en/blog/</link>
    <description>Cloud, DevOps, SRE and AI: what changes, what I make of it and how to get ready.</description>
    <language>en</language>
    <copyright>© 2026 Anthony Carbonnier</copyright>
    <atom:link href="https://metanaps.com/en/blog/feed.xml" rel="self" type="application/rss+xml"></atom:link>
    <lastBuildDate>Thu, 19 Nov 2026 19:50:53 +0000</lastBuildDate>
    <image>
      <url>https://metanaps.com/assets/favicon-192.png</url>
      <title>Blog · Metanaps — Anthony Carbonnier</title>
      <link>https://metanaps.com/en/blog/</link>
      <width>144</width>
      <height>144</height>
    </image>
    <item>
      <title>Kiri: Your Containers Without Docker Desktop, Really That Much Faster?</title>
      <link>https://metanaps.com/en/blog/kiri-lancez-vos-conteneurs-sans-docker-desktop/</link>
      <guid isPermaLink="true">https://metanaps.com/en/blog/kiri-lancez-vos-conteneurs-sans-docker-desktop/</guid>
      <pubDate>Thu, 19 Nov 2026 19:49:34 +0000</pubDate>
      <dc:creator>Anthony Carbonnier</dc:creator>
      <category>DevOps</category>
      <description>Kiri promises to launch containers at full speed, without Docker Desktop and without changing your habits. The project is already attracting a lot of attention, but its numbers and maturity are worth a closer look before you trust it with your work.</description>
      <content:encoded><![CDATA[<p>You open Docker Desktop to launch a container, then wait for the machine to wake up. Kiri wants to save you that little ritual, with a runtime written in Rust that runs your Docker images on macOS, Windows, and Linux.</p>
<h2 id="docker-but-with-the-commands-you-already-know">Docker, but with the commands you already know</h2>
<p>Kiri targets Docker Desktop and OrbStack users. It speaks the Docker Engine API, the interface that lets Docker CLI and other tools manage containers. <mark>Compose and buildx are also said to be compatible.</mark> So you can keep using your commands, at least in theory: the README doesn’t detail possible differences in behavior.</p>
<p><img src="https://metanaps.com/media/ae78a2814ee2f58a.webp" alt="Kiri&#39;s home page: “The modern container runtime”, 8× faster than Docker, with a window listing the web, db, api and redis containers." width="1280" height="800" loading="lazy" decoding="async"></p>
<p><em>Image: screenshot of <a href="https://runkiri.com" target="_blank" rel="noopener">runkiri.com</a>, taken on November 19, 2026</em></p>
<p>On macOS and Windows, Kiri runs containers in a small Linux virtual machine. On Linux, it can use the kernel directly, without going through that VM.</p>
<h2 id="milliseconds-that-make-you-want-to-try-it">Milliseconds that make you want to try it</h2>
<p>The README claims <mark>startup in about 2.5 ms via the Docker API</mark>, around 80 times faster than Docker Desktop. It also describes a VM that goes to sleep after 60 seconds of inactivity, a daemon of about 15 MB, and <mark>a wake-up time of 0.15 seconds</mark>. Tempting enough to get out the stopwatch... except these figures come from the project, not an independent benchmark.</p>
<p>Kiri adds local domains under <code>.kiri.local</code>, with HTTPS provided by its own certificate authority. The README also mentions a Kubernetes cluster you can start with one command and Linux machines. The options, limitations, and supported platforms aren’t specified here. Best to check before making it part of your development environment.</p>
<h2 id="stars-and-already-190-issues">Stars, and already 190 issues</h2>
<p>The <a href="https://github.com/libracoder/kiri" target="_blank" rel="noopener">libracoder/kiri repo on GitHub</a> was created on November 2, 2026, and got a commit on November 18. As of November 18, 2026, <mark>it has 10,590 stars, 475 forks, and 190 open issues.</mark> The owner is a user, the code is written in Rust, and the listed license is Apache-2.0.</p>
<p>Fifteen days old and already getting this much attention? Impressive. But stars don’t guarantee stability or quality, and those 190 open issues are worth a look before adopting the project.</p>
<h2 id="my-take-one-to-watch">My take: one to watch</h2>
<p>To try it out, head to the <a href="https://runkiri.com" target="_blank" rel="noopener">Kiri site</a> or its repo, then read the installation instructions before running anything. The sources provided don’t include a command to copy and paste here. Check the license and test in a disposable environment, with data that doesn’t matter. If you’re comparing options for working with AWS locally, <a href="https://metanaps.com/blog/aws-emulator-testez-vos-applications-aws-en-local-sans-deploiement/">this article on AWS Emulator</a> covers another part of the same problem.</p>
<p>In my view, Kiri is worth trying locally if Docker Desktop gets on your nerves and you like keeping an eye on young projects. I wouldn’t trust it with an important workflow just on the strength of its stars: the README claims impressive performance, without giving the details needed to judge how far those claims go ^^.</p>
]]></content:encoded>
    </item>
    <item>
      <title>TokenEconomizer: Can Your Coding Agents Really Cost Less?</title>
      <link>https://metanaps.com/en/blog/reduire-les-couts-d-agents-de-code-avec-tokeneconomizer/</link>
      <guid isPermaLink="true">https://metanaps.com/en/blog/reduire-les-couts-d-agents-de-code-avec-tokeneconomizer/</guid>
      <pubDate>Thu, 19 Nov 2026 16:00:24 +0000</pubDate>
      <dc:creator>Anthony Carbonnier</dc:creator>
      <category>Applied AI</category>
      <description>TokenEconomizer learns from agents’ discoveries in a repository so they don’t start from scratch every session. The reported savings are tempting, but installation and privacy deserve a look before you try it.</description>
      <content:encoded><![CDATA[<p>An agent rediscovering the same test command every session gets expensive fast. <a href="https://github.com/microsoft/TokenEconomizer" target="_blank" rel="noopener">TokenEconomizer</a> aims to reuse what it’s learned in a repository, without modifying the agent. I like the idea, but the numbers don’t replace taking a look at the installation and the data it examines.</p>
<h2 id="first-you-need-something-worth-remembering">First, you need something worth remembering</h2>
<p>To get started, choose a repository your agent uses regularly and has already worked in several times. TokenEconomizer can remember a working test command or a variable an tool needs. In a brand-new repository, it may not have much to keep in memory.</p>
<p>The project describes two mechanisms: micro-skills, which remember confirmed facts, and unused-tool pruning, in other words, removing what isn’t useful. The sources don’t specify which files or settings to change, so there’s no point in making up a configuration.</p>
<h2 id="installation-no-magic-command">Installation: no magic command</h2>
<p>Before running anything, check the documentation for prerequisites, supported agents, and installation steps. Also check how to connect the tool to your agent without modifying it. The provided sources don’t include an installation command or a version to pin.</p>
<p><img src="https://metanaps.com/media/5c19a57a815dea55.webp" alt="The microsoft/TokenEconomizer GitHub repository: the folder list, the “Experience-Driven Black-Box Cost Control for Coding Agents” description and over 7,000 stars." width="1280" height="800" loading="lazy" decoding="async"></p>
<p><em>Image: screenshot of <a href="https://github.com/microsoft/TokenEconomizer" target="_blank" rel="noopener">github.com</a>, taken on November 19, 2026</em></p>
<p>To assess the results, compare session costs before and after enabling it, using comparable tasks in the same repository. Also check whether the tasks succeed: saving tokens while failing more often is a lot less appealing.</p>
<h2 id="the-numbers-look-promising-with-one-caveat">The numbers look promising, with one caveat</h2>
<p>The project reports results on SWE-Bench, a set of software development tasks. <mark>Average cost falls by 18.4 to 30.9%</mark> across five agent-model combinations, and <mark>cost goes down in all 30 repository-combination cells.</mark> Pass@1, meaning success on the first try, <mark>increases in four out of five combinations.</mark> These figures come from the project; they don’t guarantee the same results in your repository.</p>
<p>Micro-skills are only suggested after a fact has been confirmed across three sessions. If later sessions contradict it, they can be removed too. That keeps you from carving an instruction based on a one-off incident in stone, but learning starts more slowly.</p>
<p>One more thing to check: the project says a model reviews recent sessions every three runs to suggest facts. The documentation should explain what data this analysis uses and how it fits with your privacy rules. The sources don’t detail data handling or deployment options. I’d look into that before connecting the tool, not after.</p>
<p>As of November 18, 2026, the GitHub repository showed 7,619 stars, 520 forks, and 32 open issues. That’s a snapshot, not a measure of quality or maintenance.</p>
<h2 id="my-take-measure-before-you-get-carried-away">My take: measure before you get carried away</h2>
<p>The same caution applies to rolling back: the sources don’t say how to disable TokenEconomizer, remove its data, or restore the previous configuration. Check the uninstall procedure before trying it, then follow it if you remove the tool.</p>
<p>I find the approach interesting for repositories where agents keep repeating the same discoveries. The reported gains are worth a measured trial, as long as you check installation and privacy first. Then compare cost and success on a representative repository, without confusing a promise with results in your own setup.</p>
]]></content:encoded>
    </item>
    <item>
      <title>AWS Emulator: Test Your AWS APIs Without Waiting for Deployment</title>
      <link>https://metanaps.com/en/blog/aws-emulator-testez-vos-applications-aws-en-local-sans-deploiement/</link>
      <guid isPermaLink="true">https://metanaps.com/en/blog/aws-emulator-testez-vos-applications-aws-en-local-sans-deploiement/</guid>
      <pubDate>Tue, 17 Nov 2026 00:16:02 +0000</pubDate>
      <dc:creator>Anthony Carbonnier</dc:creator>
      <category>DevOps</category>
      <description>An agent spent four nights rebuilding a service, then about two hours with a specification and a test environment. AWS Emulator lets you check interactions with AWS APIs locally, without waiting for a cloud deployment.</description>
      <content:encoded><![CDATA[<p>An agent spent four nights rebuilding a service. With a specification and a test environment, it took <mark>about two hours</mark>. According to AWS, the difference comes down to one detail that changes everything: the agent could finally check its work. AWS Emulator lets you test interactions with AWS services locally, without deploying to the cloud.</p>
<h2 id="test-aws-without-leaving-your-machine">Test AWS without leaving your machine</h2>
<p>The emulator runs on your machine and exposes a local HTTPS endpoint. You can send it requests using the AWS CLI, SDKs, or coding agents, with the usual AWS APIs. Your existing applications and tests can stay as they are. For example, a test that creates a DynamoDB table, writes an item to it, then reads it back can also run locally.</p>
<p>AWS relies on its public API models, the same ones used by the SDKs. The company says this means requests and errors have the same shape as they do on AWS. The emulator supports APIs for services with a model in the public API models repository. The New Stack mentions <mark>more than 200 services</mark>. That's quite a crowd.</p>
<h2 id="there-s-one-little-catch-with-the-installation">There's one little catch with the installation</h2>
<p>On Linux, you need an RPM or Debian package. On macOS or Windows, you need a Docker container. <mark>AWS doesn't provide a ready-to-use Docker image.</mark> At startup, the emulator generates credentials, a configuration file, and a certificate in <code>~/.aws/emulator/9297/</code>. You then configure your tools to point to the local endpoint.</p>
<p>AWS says there are no additional service fees. Resources are still billed when you deploy them on AWS. In short, local testing saves you from waiting for deployment, not from the bill once you're in production...</p>
<h2 id="an-agent-needs-to-be-able-to-correct-itself">An agent needs to be able to correct itself</h2>
<p>A coding agent works through successive attempts: it writes, runs, checks the result, and fixes things. If its environment doesn't reproduce the behavior of the services it calls, it can mistake an error for success, or vice versa. Testing AWS APIs locally gives it feedback without waiting for a deployment to a region. Much better than discovering the problem afterward!</p>
<p>Swami Sivasubramanian's account offers a striking example. Building the first version of a replication engine would have taken about four months. Later, he asked Kiro to rebuild it. After four nights of supervision, he realized the agent didn't know how to check its results. Once given a specification and a test environment, the rebuild took about two hours. Not every project will move that fast, of course, but the example shows why an agent needs to tell a correct change from an error.</p>
<p>The New Stack reports that autonomous agents wrote the emulator using AWS's public models. Each behavior is checked with a fidelity test, run first against real AWS and then against the emulator. Security agents also review the code during a validation step. In my view, these checks tell us more than the mere fact that an agent wrote the code: they explain how AWS checks the result.</p>
<h2 id="my-take-a-testing-tool-not-a-green-light">My take: a testing tool, not a green light</h2>
<p>The emulator covers modeled APIs, not the internal services your team has built around them. An application can pass its tests against the emulator and still hit a defect in a homegrown dependency or in a real environment. API fidelity helps, but it doesn't replace testing your own systems or pre-production checks.</p>
<p>It remains to be seen how it will perform in projects that go beyond the supported APIs, and how much confidence teams will place in its results. The fidelity tests AWS describes check behaviors; they don't guarantee that every business scenario is covered.</p>
<p>To me, AWS Emulator is a step in the right direction: the potential gain comes from fast, credible feedback, not simply adding an agent to development. I'd treat it as an additional test environment, with scenarios for important integrations and separate checks for internal services. If you use agents, the key question is whether they can see their mistakes and fix them before deployment.</p>
]]></content:encoded>
    </item>
    <item>
      <title>ARTEX: FBI warns of AI-powered automated attacks</title>
      <link>https://metanaps.com/en/blog/artex-automatise-les-cyberattaques-le-fbi-alerte-sur-la-menace/</link>
      <guid isPermaLink="true">https://metanaps.com/en/blog/artex-automatise-les-cyberattaques-le-fbi-alerte-sur-la-menace/</guid>
      <pubDate>Mon, 16 Nov 2026 13:00:16 +0000</pubDate>
      <dc:creator>Anthony Carbonnier</dc:creator>
      <category>Cloud security</category>
      <description>ARTEX agents stole credentials, tried to bypass KYC checks with deepfakes, and took part in campaigns described by Amazon. Cloudflare’s figures show how quickly these tools are spreading.</description>
      <content:encoded><![CDATA[<p>AI agents powered by ARTEX carried out offensive campaigns with little human involvement. Credential theft, real-time deepfakes to bypass KYC checks, a ConnectWise compromise that affected Microsoft 365 customer environments: the FBI has plenty to get our attention. Not exactly the kind of automation we want to see spreading!</p>
<h2 id="artex-strings-tasks-together">ARTEX strings tasks together</h2>
<p>ARTEX is an open-source, agent-powered penetration testing framework. In plain English, it coordinates reconnaissance and exploitation steps, and can connect to existing security tools, including through MCP. Operators give it goals and constraints in persistent instructions. The agents can then identify targets or check exposed credentials.</p>
<p><img src="https://metanaps.com/media/5a1ea683536e71cc.webp" alt="ARTEX coordinates agents by a server; a shield signals an alert about stolen keys." width="1185" height="740" loading="lazy" decoding="async"></p>
<p><em>Image: AI-generated drawing for this passage</em></p>
<p>Amazon Threat Intelligence uses MadPot, a network of decoy services that logs attackers’ requests. Between October 3 and November 5, 2026, ARTEX agents sent these sensors requests containing traces of their own campaigns. Amazon assessed activity that potentially targeted some AWS customers and notified them when necessary. The company couldn’t confirm the attacks against other organizations.</p>
<p>In the campaign described by Amazon, the stolen keys were used to pay for access to AI models used to keep searching for credentials. Another operation targeted cryptocurrency exchanges’ KYC checks, using real-time deepfake injection. So we’re talking about campaigns with distinct objectives, not just a technical demo.</p>
<h2 id="frameworks-are-spreading-fast">Frameworks are spreading fast</h2>
<p>Cloudflare reports a sharp rise in penetration testing and orchestration tools built on large language models. In three days, its systems saw more than 50 framework families trying to scan its customers. The IP addresses associated with these tools number in the thousands, and the daily volume of requests targeting AI-powered pentesting and exploitation tools has reportedly increased <mark>30-fold since October 1</mark>.</p>
<p><img src="https://metanaps.com/media/b609c3f52764c798.webp" alt="ARTEX, an attack robot, targets a server as a shield blocks access, labeled “ACCÈS BLOQUÉ”." width="1280" height="800" loading="lazy" decoding="async"></p>
<p><em>Image: AI-generated drawing for this passage</em></p>
<p>ARTEX’s creator took it down on October 8 and condemned its malicious use. But forks, copies derived from the project, and similar frameworks have multiplied. Cloudflare cites more than <mark>12,000 repositories</mark>, including several hundred forks of the original repository. Taking down the original project didn’t stop its descendants.</p>
<p>For security teams, that changes the scale of what needs monitoring. Agents can map environments, test access, and chain actions together at scale. Cloudflare also reports the use of stolen customer credentials to move laterally, in other words, from one system to another. So it’s better to monitor behavior and access, not just look for the name of a known tool.</p>
<h2 id="my-take-start-with-access">My take: start with access</h2>
<p>Cloudflare recommends rate-limiting rules in the WAF, anomaly detection on web applications, bot protection, identity checks for sensitive applications, strict schema validation, and threat intelligence. The goal: spot automated requests, limit access to exposed applications, and reduce opportunities for abuse.</p>
<p>We need to keep a cool head when it comes to attribution. Amazon links the campaigns to ARTEX with moderate confidence, based on references to the tool and technical indicators. Attribution to China-linked groups is also assessed with moderate confidence. Honeypot data sheds light on the methods and intentions recorded by the agents, but doesn’t confirm the impact of every operation on the organizations targeted.</p>
<p>To me, the signal worth taking seriously is the ability to chain offensive tasks together with little supervision. The original repository is gone, but forks and comparable tools are already circulating... I’d start by checking exposed access, identity controls for sensitive applications, and detection of automated behavior. And above all, by reducing what a stolen credential can reach. That’s already plenty of concrete work.</p>
]]></content:encoded>
    </item>
  </channel>
</rss>
