Cathedral is building CPU sandbox infrastructure for reinforcement learning, AI agents, and evaluation on Bittensor SN94. Here’s how it works and why it matters. […]
<p class="wp-block-paragraph">Open models have made capable AI more accessible to independent researchers and smaller teams, but a capable model is only the starting point. </p> <p class="wp-block-paragraph">Coding agents and other tool-using systems need to write and execute code, call tools, run tests, and receive reliable feedback on whether they actually completed a task.</p> <p class="wp-block-paragraph">If reinforcement learning requires thousands of attempts, those attempts need somewhere to run safely and consistently, without allowing the model to interfere with the system evaluating it.</p> <p class="wp-block-paragraph">That is the problem Cathedral is building around Bittensor, the project now operates as SN94, providing CPU-based execution environments for reinforcement learning, evaluation, and AI agents.</p> <figure class="wp-block-embed is-type-rich is-provider-x wp-block-embed-x"><div class="wp-block-embed__wrapper"> <div class="embed-x"><blockquote class="twitter-tweet" data-width="500" data-dnt="true"><p lang="en" dir="ltr">Cathedral is now on SN94.<br><br>Open models are within reach. What coding and tool-using RL still needs is somewhere for the code to run: fast, isolated CPU sandboxes.<br><br>That's the layer we're building, for open teams and for Bittensor. Live in limited beta. <a href="https://t.co/vWYo1IOz12">https://t.co/vWYo1IOz12</a></p>— Cathedral (@cathedralhq) <a href="https://x.com/cathedralhq/status/2104351337823473784?ref_src=twsrc%5Etfw">September 27, 2026</a></blockquote><script async src="https://platform.x.com/widgets.js" charset="utf-8"></script></div> </div></figure> <h2 class="wp-block-heading">What Is Cathedral?</h2> <p class="wp-block-paragraph">Cathedral is building execution infrastructure that provides isolated Linux environments where AI-generated code and agent workloads can run, giving teams a place to execute tasks, run evaluations, and generate the feedback needed for reinforcement learning. <a href="https://cathedral.computer/docs/">Cathedral Docs</a></p> <p class="wp-block-paragraph">A model producing code that looks correct does not mean the code works. It needs to be executed and tested, while the evaluation environment needs enough separation to prevent the model from manipulating the conditions under which it is judged.</p> <p class="wp-block-paragraph">At scale, that becomes a serious infrastructure problem where teams have to manage machines, sandboxing, environment setup, failures, and large numbers of repeated executions. Cathedral’s approach is to provide that execution layer instead of making every team build it from scratch.</p> <h2 class="wp-block-heading">Why Reinforcement Learning Needs an Execution Layer</h2> <figure class="wp-block-image size-full is-resized"><img loading="lazy" decoding="async" width="680" height="383" src="https://taodaily.io/wp-content/uploads/2026/09/image-285.png" alt="" class="wp-image-25118" style="aspect-ratio:2.1591695501730106;width:624px;height:auto" srcset="https://taodaily.io/wp-content/uploads/2026/09/image-285.png 680w, https://taodaily.io/wp-content/uploads/2026/09/image-285-300x169.png 300w, https://taodaily.io/wp-content/uploads/2026/09/image-285-678x381.png 678w" sizes="(max-width: 680px) 100vw, 680px" /><figcaption class="wp-element-caption"><a href="https://pbs.twimg.com/media/HTQmS9mX0AAhSs0?format=png&name=small">How RL Works On Execution</a></figcaption></figure> <p class="wp-block-paragraph">Reinforcement learning depends on repeated interaction. A model attempts a task, the environment records what happened, and that feedback helps determine what the model learns.</p> <p class="wp-block-paragraph">For a coding agent, that can mean generating code, executing it, running tests, calling tools, and checking the result. None of this works without an environment capable of safely handling those actions.</p> <p class="wp-block-paragraph">The evaluation system also has to remain separate from the model. If an agent can access or alter its grader, it could optimize for the evaluation mechanism instead of solving the task.</p> <p class="wp-block-paragraph">That makes isolated execution part of the training infrastructure itself. </p> <h2 class="wp-block-heading">Cathedral SN94’s Architecture</h2> <p class="wp-block-paragraph">Cathedral uses Firecracker micro-VMs to create isolated Linux sandboxes. </p> <ul class="wp-block-list"> <li>Its Standard box is a dedicated machine capable of hosting up to 13 isolated sandboxes and currently costs $0.99 per box-hour. At full capacity, that works out to roughly $0.076 per sandbox-hour.</li> </ul> <ul class="wp-block-list"> <li>The system also supports snapshots and forks, allowing a team to prepare an environment once and use that state as the starting point for subsequent executions.</li> </ul> <ul class="wp-block-list"> <li>Cathedral has also integrated with existing agent tooling, including Harbor and Prime Intellect’s verifiers library. <a href="https://github.com/cathedralai/cathedral">Cathedral GitHub</a> | <a href="https://www.tbench.ai/news/announcement-2-0">Harbor</a> | <a href="https://github.com/PrimeIntellect-ai/verifiers">Prime Intellect Verifiers</a></li> </ul> <p class="wp-block-paragraph">The model produces the action, the sandbox executes it, and the evaluation system judges the result without giving the model control over the infrastructure.</p> <h2 class="wp-block-heading">Why Cathedral Moved to SN94</h2> <p class="wp-block-paragraph">Cathedral previously operated on another Bittensor subnet before announcing its move to <strong>netuid 94</strong>. Its official announcement frames SN94 around reinforcement learning execution and the infrastructure needed to support it.</p> <p class="wp-block-paragraph">The longer-term goal is to coordinate independent compute providers around a shared execution service rather than relying entirely on infrastructure operated by a single provider.</p> <figure class="wp-block-embed is-type-rich is-provider-x wp-block-embed-x"><div class="wp-block-embed__wrapper"> <div class="embed-x"><blockquote class="twitter-tweet" data-width="500" data-dnt="true"><p lang="en" dir="ltr">Cathedral is now on SN94.<br><br>Open models are within reach. What coding and tool-using RL still needs is somewhere for the code to run: fast, isolated CPU sandboxes.<br><br>That's the layer we're building, for open teams and for Bittensor. Live in limited beta. <a href="https://t.co/vWYo1IOz12">https://t.co/vWYo1IOz12</a></p>— Cathedral (@cathedralhq) <a href="https://x.com/cathedralhq/status/2104351337823473784?ref_src=twsrc%5Etfw">September 27, 2026</a></blockquote><script async src="https://platform.x.com/widgets.js" charset="utf-8"></script></div> </div></figure> <p class="wp-block-paragraph">That fits a growing need inside Bittensor where subnets that evaluate executable code or agents need somewhere to run those workloads, while reinforcement learning systems need environments where models can repeatedly act and receive feedback.</p> <h2 class="wp-block-heading">The Demand Is Already Visible</h2> <figure class="wp-block-image size-large is-resized"><img loading="lazy" decoding="async" width="1024" height="576" src="https://taodaily.io/wp-content/uploads/2026/09/image-286-1024x576.png" alt="" class="wp-image-25119" style="aspect-ratio:1.7777777777777777;width:624px;height:auto" srcset="https://taodaily.io/wp-content/uploads/2026/09/image-286-1024x576.png 1024w, https://taodaily.io/wp-content/uploads/2026/09/image-286-300x169.png 300w, https://taodaily.io/wp-content/uploads/2026/09/image-286-768x432.png 768w, https://taodaily.io/wp-content/uploads/2026/09/image-286-678x381.png 678w, https://taodaily.io/wp-content/uploads/2026/09/image-286.png 2048w" sizes="(max-width: 1024px) 100vw, 1024px" /><figcaption class="wp-element-caption">The Scale Of The Frontier</