<?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: Akhil Anand</title>
    <description>The latest articles on DEV Community by Akhil Anand (@akhil_5d07).</description>
    <link>https://dev.to/akhil_5d07</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%2F4240629%2F04fce61f-b7ea-4621-8a38-5173eef51cad.png</url>
      <title>DEV Community: Akhil Anand</title>
      <link>https://dev.to/akhil_5d07</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/akhil_5d07"/>
    <language>en</language>
    <item>
      <title>How to stop a runaway AI agent loop before it runs up your OpenAI or Anthropic bill</title>
      <dc:creator>Akhil Anand</dc:creator>
      <pubDate>Thu, 12 Nov 2026 09:03:39 +0000</pubDate>
      <link>https://dev.to/akhil_5d07/how-to-stop-a-runaway-ai-agent-loop-before-it-runs-up-your-openai-or-anthropic-bill-58oi</link>
      <guid>https://dev.to/akhil_5d07/how-to-stop-a-runaway-ai-agent-loop-before-it-runs-up-your-openai-or-anthropic-bill-58oi</guid>
      <description>&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%2Fbeakgrms031ixujof9f4.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%2Fbeakgrms031ixujof9f4.png" alt=" " width="800" height="336"&gt;&lt;/a&gt;&lt;strong&gt;TL;DR:&lt;/strong&gt; A runaway agent loop shows up in your API traffic before it shows up on your bill. It looks like the same call repeated fast, or a context that keeps doubling. Tag every LLM call with the product feature it belongs to. Watch the last few calls per feature. When they look like a loop, throttle that one feature first and block it if the loop continues. This post has a small TypeScript circuit breaker you can copy, the limits of doing it yourself, and how the existing tools (LiteLLM, Helicone, Portkey, Langfuse and AtlasBurn) handle it.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Disclosure: I build &lt;a href="https://atlasburn.com" rel="noopener noreferrer"&gt;AtlasBurn&lt;/a&gt;, one of the tools compared below. The DIY code works without it.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  Why provider spend limits don't save you
&lt;/h2&gt;

&lt;p&gt;OpenAI and Anthropic let you set a monthly budget or alert on the whole account. That's a last line of defence, not loop protection:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It's account-wide.&lt;/strong&gt; When it trips, every AI feature in your app stops, including the ones that were fine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It's slow.&lt;/strong&gt; Usage dashboards update with a delay. An agent making several calls a second can do real damage before a limit fires.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It doesn't know what a loop is.&lt;/strong&gt; A busy, healthy day and a stuck agent look the same to a budget.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;What you want is a &lt;strong&gt;per-feature circuit breaker&lt;/strong&gt;: something that sees &lt;em&gt;which&lt;/em&gt; feature is misbehaving and stops only that feature, while the rest of the app keeps running.&lt;/p&gt;

&lt;h2&gt;
  
  
  What a runaway loop looks like in the traffic
&lt;/h2&gt;

&lt;p&gt;Two patterns cover most incidents:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Retry storm / identical calls.&lt;/strong&gt; The same model gets the same-sized input several times in a few seconds. Usually a tool call that keeps failing and an agent that keeps retrying.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Runaway context growth.&lt;/strong&gt; Each call carries the previous output back in, so input tokens climb steadily (often doubling) across a burst of calls. The cost per call grows with it.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A third signal, a &lt;strong&gt;request-rate burst&lt;/strong&gt;, catches fan-out bugs: a job that suddenly sends many requests per second.&lt;/p&gt;

&lt;p&gt;The catch is false positives. A normal multi-turn chat also grows its context every turn. So you need two filters: &lt;strong&gt;speed&lt;/strong&gt; (loops are fast, e.g. 4 calls inside 30 seconds) and &lt;strong&gt;size&lt;/strong&gt; (ignore growth until the context is large enough to cost real money, e.g. 4,000+ input tokens).&lt;/p&gt;

&lt;h2&gt;
  
  
  A per-feature circuit breaker in ~40 lines (TypeScript)
&lt;/h2&gt;

&lt;p&gt;This keeps the last few calls per feature in memory and pauses a feature when its recent calls look like a loop.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="nc"&gt;Call&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;at&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="n"&gt;inputTokens&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;number&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;WINDOW&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;4&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;            &lt;span class="c1"&gt;// look at the last 4 calls of a feature&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;MAX_SPAN_MS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;30_000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c1"&gt;// ...that happened within 30 seconds&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;MIN_TOKENS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;4_000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;    &lt;span class="c1"&gt;// ignore context growth below this size&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;PAUSE_MS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;10&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;60_000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;recent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Map&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Call&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;pausedUntil&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Map&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kt"&gt;number&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;isPaused&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;boolean&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pausedUntil&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;now&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;recordCall&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;call&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Call&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;void&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;calls&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;recent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;call&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;slice&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="no"&gt;WINDOW&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="n"&gt;recent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;calls&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;calls&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="no"&gt;WINDOW&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&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;calls&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="no"&gt;WINDOW&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;at&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;calls&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;at&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="no"&gt;MAX_SPAN_MS&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// not loop-speed&lt;/span&gt;

  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;sameModel&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;calls&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;every&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="n"&gt;calls&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;calls&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;map&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;inputTokens&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;identical&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;sameModel&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;every&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;t&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;growing&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt;
    &lt;span class="n"&gt;sameModel&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
    &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;every&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="n"&gt;t&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
    &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
    &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="no"&gt;WINDOW&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
    &lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="no"&gt;WINDOW&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="no"&gt;MIN_TOKENS&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;identical&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="n"&gt;growing&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;pausedUntil&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;now&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="no"&gt;PAUSE_MS&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Wrap your LLM calls with it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s2"&gt;"openai"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;isPaused&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;recordCall&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s2"&gt;"./loop-breaker"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;params&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;ChatCompletionCreateParamsNonStreaming&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="nf"&gt;isPaused&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="c1"&gt;// Fail gracefully for this feature only. Everything else keeps working.&lt;/span&gt;
    &lt;span class="k"&gt;throw&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Error&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sb"&gt;`AI feature "${feature}" is paused: loop detected`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;res&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;params&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nf"&gt;recordCall&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;at&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;now&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;params&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;inputTokens&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;res&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;usage&lt;/span&gt;&lt;span class="o"&gt;?&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;prompt_tokens&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;res&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"support-chat"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"gpt-4o-mini"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;role&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"user"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;content&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Hi"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's enough to stop the classic "agent retried a failing tool 400 times overnight" incident in a single process.&lt;/p&gt;

&lt;h3&gt;
  
  
  Where the DIY version falls short
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;One process only.&lt;/strong&gt; With several instances or serverless functions, each one sees a fraction of the calls. You need shared state such as Redis.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It pauses immediately.&lt;/strong&gt; Real systems want a ladder: &lt;em&gt;throttle first&lt;/em&gt; (space calls out), then block only if the pattern keeps going or crosses a dollar amount. Otherwise one false positive takes a feature down.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No budgets.&lt;/strong&gt; It doesn't know what a call costs, so it can't stop a feature that is expensive but not looping.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Streaming.&lt;/strong&gt; With streaming responses you only get token usage if you ask for it (&lt;code&gt;stream_options: { include_usage: true }&lt;/code&gt; on OpenAI) and read the final chunk.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every provider and SDK.&lt;/strong&gt; You have to wrap each client (OpenAI, Anthropic, Gemini…) and each code path that calls them.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How existing tools handle it
&lt;/h2&gt;

&lt;p&gt;From each project's public docs as of November 2026. Features move fast, so check their docs before deciding.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tool&lt;/th&gt;
&lt;th&gt;Cost per feature&lt;/th&gt;
&lt;th&gt;Stops spend automatically&lt;/th&gt;
&lt;th&gt;Loop / retry-storm detection&lt;/th&gt;
&lt;th&gt;How it's added&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;LiteLLM&lt;/strong&gt; (proxy)&lt;/td&gt;
&lt;td&gt;By key, team, user or tag&lt;/td&gt;
&lt;td&gt;Yes: hard budgets and rate limits per key, team, user or tag (needs a Postgres database)&lt;/td&gt;
&lt;td&gt;Agent session caps (max iterations, budget per session) for agents on its Agent Gateway; no repeated-call or context-growth detector&lt;/td&gt;
&lt;td&gt;Self-hosted or managed gateway&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Helicone&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Via custom properties (a &lt;code&gt;Helicone-Property-*&lt;/code&gt; header)&lt;/td&gt;
&lt;td&gt;Rate limits, including cost-based ones per user or property (returns 429)&lt;/td&gt;
&lt;td&gt;Not a dedicated loop detector&lt;/td&gt;
&lt;td&gt;Gateway / base-URL change&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Portkey&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Via metadata on requests&lt;/td&gt;
&lt;td&gt;Budget and rate limits on keys (Enterprise and select Pro plans)&lt;/td&gt;
&lt;td&gt;Not a dedicated loop detector&lt;/td&gt;
&lt;td&gt;Gateway / base-URL change&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Langfuse&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Via tags and traces&lt;/td&gt;
&lt;td&gt;No blocking; cost-threshold alerts (Slack, webhook…)&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;td&gt;SDK / OpenTelemetry tracing&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;AtlasBurn&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Per feature via a request header&lt;/td&gt;
&lt;td&gt;Yes: per-feature throttle and kill switch, plus daily/hourly budgets&lt;/td&gt;
&lt;td&gt;Yes: identical-call loops, runaway context growth, request-rate bursts&lt;/td&gt;
&lt;td&gt;Node.js SDK that patches &lt;code&gt;fetch&lt;/code&gt;, or a base-URL proxy for any language&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Rule of thumb: if you already run a gateway and mostly need &lt;strong&gt;budgets&lt;/strong&gt; (or a hard cap on agent iterations), LiteLLM or Portkey cover it. If you need &lt;strong&gt;tracing and evals&lt;/strong&gt;, Langfuse. If the specific worry is &lt;strong&gt;an agent looping in one feature and you want only that feature stopped&lt;/strong&gt;, that's the case AtlasBurn is built for.&lt;/p&gt;

&lt;h2&gt;
  
  
  The same breaker without writing it: AtlasBurn
&lt;/h2&gt;

&lt;p&gt;AtlasBurn runs the checks above server-side for every feature, across instances:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Identical-call loops:&lt;/strong&gt; the last 4 calls of a feature use the same model with the same input size, within 30 seconds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Runaway context growth:&lt;/strong&gt; input tokens rise on every one of the last 4 calls (same model, within 30 seconds), at least double across them, and the latest call is 4,000+ tokens. That last condition keeps normal chats from tripping it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Request-rate bursts:&lt;/strong&gt; sustained bursts of 10+ requests per second.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Budgets:&lt;/strong&gt; daily and hourly limits.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Loops are &lt;strong&gt;throttled first&lt;/strong&gt; (calls spaced 2 seconds apart) and only &lt;strong&gt;blocked&lt;/strong&gt; if the loop is still going 60 seconds later, or once it has cost 2% of your daily budget (minimum $0.25). Budget breaches block straight away. Modes run from Alert (never blocks) to Throttle to Kill Switch, and only the looping feature is affected.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Node.js / TypeScript:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm install @atlasburn/sdk
npx abn init
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;initAtlasBurnAuto&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;atlasburnFeature&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s2"&gt;"@atlasburn/sdk"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="nf"&gt;initAtlasBurnAuto&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;apiKey&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;env&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;ATLASBURN_KEY&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="c1"&gt;// Tag calls with the feature they belong to (the tag is stripped before the provider sees it).&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;params&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;headers&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;atlasburnFeature&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"support-chat"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A blocked feature gets a provider-shaped &lt;code&gt;429&lt;/code&gt; with &lt;code&gt;x-atlasburn-blocked: 1&lt;/code&gt; and &lt;code&gt;x-should-retry: false&lt;/code&gt;, so well-behaved SDKs stop retrying. If AtlasBurn can't be reached, calls go through as normal (it fails open).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Python or any other language:&lt;/strong&gt; point your client at the proxy and send the feature as a header:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"https://proxy.atlasburn.com/openai/v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;default_headers&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"AtlasBurn-API-Key"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&amp;lt;your AtlasBurn ingest key&amp;gt;"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"gpt-4o-mini"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"role"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"user"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"content"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Hi"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;extra_headers&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"x-atlasburn-feature"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"support-chat"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It reads usage only (model, token counts, timing, your feature label), never prompts or outputs. The SDK is open source (Apache-2.0, &lt;a href="https://github.com/AtlasBurn/sdk" rel="noopener noreferrer"&gt;github.com/AtlasBurn/sdk&lt;/a&gt;), and the free plan covers up to 10M tokens a month with guardrails and the kill switch included.&lt;/p&gt;

&lt;p&gt;Docs: &lt;a href="https://docs.atlasburn.com/docs/guardrails/" rel="noopener noreferrer"&gt;Guardrails and the kill switch&lt;/a&gt; · &lt;a href="https://docs.atlasburn.com/docs/quickstart/" rel="noopener noreferrer"&gt;Quickstart&lt;/a&gt; · &lt;a href="https://docs.atlasburn.com/docs/proxy-quickstart/" rel="noopener noreferrer"&gt;Edge Proxy for Python, Go and other languages&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;How do I stop an AI agent from looping and burning API credits?&lt;/strong&gt;&lt;br&gt;
Cap iterations in the agent framework (e.g. &lt;code&gt;max_turns&lt;/code&gt;), and add a circuit breaker in the request path that detects repeated calls or fast context growth per feature and pauses that feature. Keep a provider-level spend limit as the final backstop.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I detect a retry storm in an LLM app?&lt;/strong&gt;&lt;br&gt;
Look at the last few calls of each feature. Several calls to the same model with the same input size within a few seconds is a retry storm. Block or slow that feature before the next call is sent.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I track LLM costs per feature?&lt;/strong&gt;&lt;br&gt;
Attach a feature label to every request (a header, metadata field or tag), record the token usage from each response, and multiply by the model's per-token price. Gateways and tools like LiteLLM, Helicone, Langfuse and AtlasBurn do this for you.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I block one AI feature without taking down the whole app?&lt;/strong&gt;&lt;br&gt;
Yes, if enforcement is scoped per feature rather than per account or API key. Give each feature its own budget/limit (a separate key or tag in a gateway), or use a tool that scopes its kill switch to the feature that's looping.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do OpenAI and Anthropic have per-feature spend limits?&lt;/strong&gt;&lt;br&gt;
Not per feature. Their limits and alerts apply at the organization, project or workspace level, so a loop in one feature can trip a limit that stops everything in that project.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;What's the worst agent loop you've hit? I'm collecting real incidents to tune the thresholds above, so please share in the comments.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>openai</category>
      <category>llm</category>
      <category>typescript</category>
    </item>
  </channel>
</rss>
