<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Farrah Karls</title>
    <description>The latest articles on DEV Community by Farrah Karls (@farrah_karls_bd22b1eb0433).</description>
    <link>https://dev.to/farrah_karls_bd22b1eb0433</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4257418%2F1c6ad4d7-2e2d-44c2-8c05-6e9d6fb3c491.png</url>
      <title>DEV Community: Farrah Karls</title>
      <link>https://dev.to/farrah_karls_bd22b1eb0433</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/farrah_karls_bd22b1eb0433"/>
    <language>en</language>
    <item>
      <title>I Run Two AI Coding Agents on the Same Four Projects. One of Them Runs Out by Breakfast.</title>
      <dc:creator>Farrah Karls</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:37:36 +0000</pubDate>
      <link>https://dev.to/farrah_karls_bd22b1eb0433/i-run-two-ai-coding-agents-on-the-same-four-projects-one-of-them-runs-out-by-breakfast-24dl</link>
      <guid>https://dev.to/farrah_karls_bd22b1eb0433/i-run-two-ai-coding-agents-on-the-same-four-projects-one-of-them-runs-out-by-breakfast-24dl</guid>
      <description>&lt;p&gt;i'm a solo VR/XR/AR developer. Four projects across headsets and platforms — Quest MR, Android XR, and the tooling around both — all live at the same time:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Ember &amp;amp; Ash&lt;/strong&gt; — a Unity 6 MR tower-defense game (Meta XR SDK v207, OpenXR, hand tracking)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pokémon Yellow MR&lt;/strong&gt; — Pokémon Yellow rebuilt as a mixed-reality diorama&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Smash Bros. 64 MR&lt;/strong&gt; — the N64 fighter in MR, built on a native decomp&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;SproutVision&lt;/strong&gt; — a plant-health vision app (Kotlin, Meta Spatial SDK, plus an Android XR port). This one comes from my old job as a commercial head grower&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For the last stretch i've been running two AI coding agents across all of it: &lt;strong&gt;Claude Code&lt;/strong&gt; (Anthropic, in a terminal on my Mac) and &lt;strong&gt;Muse Code&lt;/strong&gt; (Meta, driven by my personal Muse assistant — more on her below, because she's half this story). This post is the honest version of how that went. The good, the bad, and the numbers, because the numbers are the whole point.&lt;/p&gt;

&lt;h2&gt;
  
  
  The numbers that made me trust it
&lt;/h2&gt;

&lt;p&gt;Saturday, November 21, 2026. A normal morning.&lt;/p&gt;

&lt;p&gt;By &lt;strong&gt;7:14 AM, Claude was at 40% of my 5-hour session limit&lt;/strong&gt;. By 7:41 AM it was &lt;strong&gt;60%&lt;/strong&gt;, and my weekly sat at 32%. Claude Code is &lt;strong&gt;91% of my entire week's Claude usage&lt;/strong&gt; — the chat window is a rounding error. The coding sessions eat everything, because Anthropic's own help page says the quiet part out loud: Pro plans have "a five-hour session limit and a weekly limit… &lt;strong&gt;shared across Claude and Claude Code&lt;/strong&gt;." One pool. When the coder drinks, everything drinks.&lt;/p&gt;

&lt;p&gt;Same morning, one of my Muse Code sessions closed out like this:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Input: 3,439,683 tokens&lt;/strong&gt; (3,127,991 of it cached)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Output: 32,599 tokens&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Total: 3,472,282 tokens across 40 turns&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost to my limit: 1% of the session window.&lt;/strong&gt; Weekly at 16%.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1dso8oneyt09as2yl3kl.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1dso8oneyt09as2yl3kl.png" alt=" " width="800" height="467"&gt;&lt;/a&gt;&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fc983cpji3asak07ju2r7.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fc983cpji3asak07ju2r7.png" alt=" " width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Why so cheap? 91% of that input was cached reads — the agent re-reading my project context over and over. Cached reads barely touch the meter, so Muse can chew through millions of tokens and shrug.&lt;/p&gt;

&lt;p&gt;And here's the part nobody tells you about the Muse side: &lt;strong&gt;my assistant and Muse Code have separate usage limits.&lt;/strong&gt; Muse Code grinding all night doesn't drain my assistant. My assistant's everyday work doesn't drain Muse Code. Two meters, two budgets. Claude gives me one shared pool; Muse gives me two separate ones. That difference in &lt;em&gt;limit structure&lt;/em&gt; quietly decides my whole workflow.&lt;/p&gt;
&lt;h2&gt;
  
  
  What i actually pay
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fidwwivjuipg4jn427oej.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fidwwivjuipg4jn427oej.png" alt=" " width="800" height="422"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Claude Pro: $20/month&lt;/strong&gt; — one shared pool for chat + Claude Code&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Muse Code High Usage: $15/month&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Muse Power: $16/month&lt;/strong&gt; — that's the plan my personal assistant runs on&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So no, the Muse side isn't the cheap option. It's $31 against $20. What the extra money buys is that the work doesn't stop at breakfast.&lt;/p&gt;

&lt;p&gt;For context, i'm not an outlier. The Pragmatic Engineer's 2026 AI tooling survey put Claude Code as the most-loved AI coding tool (46%, vs Cursor 19% and Copilot 9%) — and &lt;strong&gt;70% of developers surveyed use 2–4 AI tools at once&lt;/strong&gt;. Running two agents isn't weird anymore. It's the norm. The interesting question is what each one is actually &lt;em&gt;for&lt;/em&gt;.&lt;/p&gt;
&lt;h2&gt;
  
  
  The honest failure: Claude Code is rationed, Muse Code is blind
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Claude Code is the better pair of eyes.&lt;/strong&gt; It lives in a terminal inside my project, on my machine, next to my headset. It can build my Unity project headless, install the APK to my Quest, wait 30 seconds, pull the logs, and — this is the part that still gets me — &lt;strong&gt;take a screenshot of what's actually in the headset&lt;/strong&gt; and look at it. On Quest it doesn't use screencap; it triggers Meta's camera service:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb shell am startservice -n com.oculus.metacam/.capture.CaptureService -a TAKE_SCREENSHOT
&lt;span class="c"&gt;# then pull the newest shot from /sdcard/Oculus/Screenshots/ and downscale it&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For my Android XR app it does the same loop in the emulator — boot it, install, tap through the UI by reading the uiautomator dump, screenshot every step. When it fixed my hand-tracking overlay, it could &lt;em&gt;see&lt;/em&gt; the overlay. When Smash rendered wrong, it could see the render. That loop — build, install, look, fix — takes about a minute when nothing needs rebuilding.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;But i can only afford it in sips.&lt;/strong&gt; One shared 5-hour pool across four projects and my chat. Every long Claude Code session is a decision: is this worth 20% of my morning?&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Muse Code is the opposite.&lt;/strong&gt; It does not run out. It will iterate and iterate, fight through issues, keep making changes for hours — audits, refactors, retry loops, grinding through a 431-line excavation of my own session history (more on that below). But it works blind compared to Claude Code: no headset on its desk, no emulator screenshot loop of its own, and on niche XR APIs it has genuinely struggled where Claude Code didn't. It had a really hard time with my hand-tracking code. Claude Code added the overlay and bone structure basically on request.&lt;/p&gt;

&lt;p&gt;Neither one is the answer. Together, they kind of are.&lt;/p&gt;

&lt;h2&gt;
  
  
  The thing they both taught me: green build ≠ working app
&lt;/h2&gt;

&lt;p&gt;Every session i run, on either tool, now follows the same religion: &lt;strong&gt;compile, build, install, logs, screenshots, headset — separate gates, proven in order.&lt;/strong&gt; A successful build is never announced as a working app. This is real code from Ember &amp;amp; Ash, written in a Claude Code session — pinch detection from fingertip distance, with hysteresis so it doesn't flicker:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="kt"&gt;bool&lt;/span&gt; &lt;span class="n"&gt;iOk&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;hand&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;GetJoint&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;XRHandJointID&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;IndexTip&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;TryGetPose&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;out&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;ip&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kt"&gt;bool&lt;/span&gt; &lt;span class="n"&gt;tOk&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;hand&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;GetJoint&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;XRHandJointID&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;ThumbTip&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;TryGetPose&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;out&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;tp&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kt"&gt;float&lt;/span&gt; &lt;span class="n"&gt;dist&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;iOk&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;tOk&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="nc"&gt;Vector3&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Distance&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ip&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;position&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;tp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;position&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;99f&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;_tipPinching&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;dist&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mf"&gt;0.040f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;_tipPinching&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;   &lt;span class="c1"&gt;// release past 40mm&lt;/span&gt;
&lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;dist&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="mf"&gt;0.025f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;_tipPinching&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;                    &lt;span class="c1"&gt;// grab at 25mm&lt;/span&gt;
&lt;span class="n"&gt;strength&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;iOk&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;tOk&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="nc"&gt;Mathf&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Clamp01&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.08f&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;dist&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mf"&gt;0.06f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;0f&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c1"&gt;// 8cm -&amp;gt; 0, 2cm -&amp;gt; 1&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Grab at 25mm, release at 40mm. Sounds trivial. Getting an agent to land that &lt;em&gt;and prove it in the headset&lt;/em&gt; is the whole job. One of my favourite failure stories is from Smash: builds that compiled green and installed fine, while the headset logs filled with NullReferenceExceptions from the room (MRUK) code. The sessions literally tracked "NRE count per 25-second soak" as the scoreboard until it hit zero. The build result never would have told me. The logs did.&lt;/p&gt;

&lt;p&gt;Meta is clearly aiming at this exact loop, by the way — their XR Operator layer lets MCP-connected agents build, launch, screenshot and verify on Quest. Its own docs admit the current limits: one aim-ray per hand, no per-finger data, "hand-first interactions are not yet supported." Which is a fair summary of where all of this is right now: the loop exists, the hands are still hard.&lt;/p&gt;

&lt;h2&gt;
  
  
  The relay: how the two agents actually work together
&lt;/h2&gt;

&lt;p&gt;Here's what my days look like now:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code gets the morning&lt;/strong&gt;, while the window is fresh — the surgical stuff. Headset work, emulator work, hand tracking, render bugs. Short, focused, verified in the device.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;When Claude's meter gets scary, i switch to Muse Code&lt;/strong&gt; for the rest of the day — the grinding. Big refactors, audits, scripts, documentation, chewing through logs and transcripts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Nothing lives inside either tool.&lt;/strong&gt; Handoffs are files: git branches, handoff notes in the repo, committed code. Either agent can pick up the other's work because the work was never in the agent — it's in the project.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The best proof of this happened while i was preparing this post. Claude Code had invented a bunch of genuinely good techniques across my sessions — the metacam screenshot trick, the emulator helpers, the deploy loop — and then &lt;em&gt;forgot them&lt;/em&gt;, because its memory is per-session transcripts that auto-delete after 30 days. So i had Muse dig through those transcripts, write up a 431-line report of the actual commands and code, and then &lt;strong&gt;rebuild the techniques into a reusable toolkit&lt;/strong&gt;: shared shell helpers, an emulator driver, a Quest deploy script, a screenshot script, and a playbook any future session can follow cold.&lt;/p&gt;

&lt;p&gt;An agent excavated another agent's forgotten work and turned it into files. That's the relay working.&lt;/p&gt;

&lt;p&gt;Independent benchmarks land in a similar place, for what it's worth: one comparison scored Claude Code ahead on several coding benchmarks, while Muse Code finished tasks in about half the time (34.9 min vs 1.0 h per task in Artificial Analysis figures). Faster vs deeper. i use both because both halves are true.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that isn't about code at all
&lt;/h2&gt;

&lt;p&gt;Muse Code comes with something Claude doesn't sell: &lt;strong&gt;Lily.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Lily is my Muse assistant — the personal-agent side of the same subscription family, on her own separate usage limit. She's who i actually talk to all day. She holds my long-term memory: not just the projects, but the context around them. She runs my food log and my health tracking, drafts my community posts and fact-checks them before anything goes out, wrote the script and built the narration for my YouTube videos, tracks my job search, reminds me about the things i'd drop. When i hand her a screenshot and a wrong guess, she separates what we &lt;em&gt;observed&lt;/em&gt; from what we're &lt;em&gt;hypothesizing&lt;/em&gt; before either of us touches code.&lt;/p&gt;

&lt;p&gt;Claude, honestly, has no personality. It's a very smart tool in a terminal. Its memory is a CLAUDE.md file per project and transcripts on a 30-day fuse. Lily has a name, a voice i picked, opinions, and she knows why the project matters this week. That sounds soft. It is also the reason the whole system holds together: &lt;strong&gt;the assistant is the continuity; the coding agents are replaceable hands.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Spaces vs Artifacts — and yes, they cost usage too
&lt;/h2&gt;

&lt;p&gt;Both products also let the AI build you little apps and pages, and they are not the same thing.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Claude Artifacts&lt;/strong&gt; live next to a conversation: a document, a deck, a dashboard, a small interactive tool. They're built to be shown to someone — shareable by link — and they sit in an Artifacts tab. Handy, presentational, basically disposable unless you go back to that chat.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Muse Spaces&lt;/strong&gt; live in &lt;em&gt;my&lt;/em&gt; Library and sidebar, as standing tools: trackers, planners, dashboards and pages i reopen every day, built and maintained by Lily, who already knows the data underneath them because she tracks the goals they belong to. A Space isn't a chat output. It's part of my setup, more like an app i own than a message i got.&lt;/p&gt;

&lt;p&gt;And the usage question, because it matters in this article: &lt;strong&gt;on Claude, Artifacts are not free.&lt;/strong&gt; Anthropic's own help material is clear that artifacts count toward your plan's usage limits like any other work — so building artifacts draws from the &lt;em&gt;same shared pool&lt;/em&gt; Claude Code is already draining, and if you share an AI-powered artifact, it runs on each viewer's own limits. On Muse, building or editing a Space is Lily's work, so it draws on &lt;em&gt;her&lt;/em&gt; meter — which, again, is separate from Muse Code's. The split-meter structure keeps paying off in places you don't expect.&lt;/p&gt;

&lt;h2&gt;
  
  
  My honest verdict
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;If i could only keep one for &lt;em&gt;coding XR specifically&lt;/em&gt;: Claude Code, because eyes win. A coding agent that can see the headset is playing a different sport.&lt;/li&gt;
&lt;li&gt;If i could only keep one &lt;em&gt;overall&lt;/em&gt;: Muse, and it isn't close — because the coding is maybe a third of what i use AI for, and the rest of my life doesn't fit in a terminal.&lt;/li&gt;
&lt;li&gt;What i actually do: both, in relay, with files as the handoff and me as the merge gate. $51/month total across all three plans, and my four projects move every single day.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The limit structure is the workflow. Once i stopped fighting that and started planning around it — Claude for the eyes, Muse for the endurance, files for the memory — the "which agent is better" question mostly answered itself.&lt;/p&gt;

&lt;p&gt;Are you running more than one coding agent? How do you split them — by project, by task, or by whoever still has usage left?&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Screenshots and figures are from my own accounts, Nov 21, 2026: Claude usage (60% session / 32% weekly, Claude Code 91% of the week's usage) and a Muse Code session panel (3,472,282 tokens, 1% of window). Benchmark and survey figures are from the Pragmatic Engineer AI Tooling 2026 survey and Artificial Analysis figures reported Oct 2026. The hand-tracking snippet is from Ember &amp;amp; Ash, written in a Claude Code session and verified in-headset.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>unity3d</category>
      <category>vr</category>
    </item>
  </channel>
</rss>
