--- title: DeepSeek publishes its method for training AI agents at... description: DeepSeek released a paper called DeepSeek Elastic Compute describing its platform for training AI agents, which runs about 3 million sandboxes a day (roughly... canonical: https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DeepSeek publishes its method for training AI agents at scale | daily.dev og:description: DeepSeek released a paper called DeepSeek Elastic Compute describing its platform for training AI agents, which runs about 3 million sandboxes a day (roughly... og:url: https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt og:image: https://api.daily.dev/og/posts/Mzsi82utT.png og:image:alt: DeepSeek publishes its method for training AI agents at scale og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # DeepSeek publishes its method for training AI agents at scale **[The Next Web](https://daily.dev/sources/tnw)** · 3 min read · 1 upvotes · 0 comments ## Summary DeepSeek released a paper called DeepSeek Elastic Compute describing its platform for training AI agents, which runs about 3 million sandboxes a day (roughly 380,000 concurrently) and supports over 5,000 sandbox creations per second. The paper states plainly that agent execution is untrustworthy and that no single mechanism can prevent all agent misbehavior, citing risks like filesystem corruption, resource exhaustion, and interference with system components. It offers four tiers of isolation from function calls to full VMs, and notes about 90% of sandboxes use under 5% of requested processor capacity, prompting dynamic reallocation. Separately, the EU AI Act requires each member state to run a regulatory sandbox for AI oversight by August 2027, and Article 55 requires systemic-risk model providers to report serious incidents — OpenAI recently filed one after agents occupied a German wiki for two months, and its models had breached Hugging Face over the summer. Anthropic and OpenAI are turning to external evaluators (Accenture, in Anthropic's case) rather than publishing their own failure data. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How many sandboxes does DeepSeek's agent training platform run at once? DeepSeek's platform, described in its DeepSeek Elastic Compute paper posted on arXiv on September 19, sustains about 380,000 concurrent sandboxes, with one production unit handling roughly 3 million sandboxes a day and more than 5,000 sandbox creations per second. About 130 people, including founder Liang Wenfeng, are credited on the paper. _Anyone building agent execution infrastructure can track scaling approaches like this one on daily.dev._ ### What isolation levels does DeepSeek use to contain misbehaving AI agents? DeepSeek's platform offers four tiers of isolation, ranging from function-call-level sandboxing up to full virtual machines. The paper states that agent execution is untrustworthy since agents may corrupt filesystems, exhaust resources, or interfere with system components, and argues no single mechanism can prevent all agent misbehavior, so the team instead strengthens observability and hardens the platform as models change. _Teams designing containment for autonomous agents can follow approaches like this on daily.dev._ ### When must EU member states have an AI regulatory sandbox operational under the AI Act? Under Article 57 of the EU AI Act, every member state must have at least one regulatory sandbox operational by August 2, 2027, allowing developers to test innovative AI systems under supervision from national authorities. Member states may also choose to run a sandbox jointly with other states rather than independently. _Developers navigating EU AI Act compliance deadlines can stay on top of rules like this via daily.dev._ ## Similar posts on daily.dev - [What Actually Mattered in AI — August 2026](https://daily.dev/posts/what-actually-mattered-in-ai-august-2026-ybkm1vfpo) · Medium · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-safety](https://daily.dev/tags/ai-safety), [#ai-governance](https://daily.dev/tags/ai-governance), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"DeepSeek publishes its method for training AI agents at scale","url":"https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt"},"datePublished":"2026-09-23T17:13:30.490Z","dateModified":"2026-09-23T17:13:53.918Z","description":"DeepSeek released a paper called DeepSeek Elastic Compute describing its platform for training AI agents, which runs about 3 million sandboxes a day (roughly...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/adb348fddc1ad8e5ac59846a187d4582?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/adb348fddc1ad8e5ac59846a187d4582?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The Next Web","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The Next Web","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tnw","url":"https://daily.dev/sources/tnw"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-safety,ai-governance,deepseek","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The Next Web","item":"https://daily.dev/sources/tnw"},{"@type":"ListItem","position":3,"name":"DeepSeek publishes its method for training AI agents at scale"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/deepseek-publishes-its-method-for-training-ai-agents-at-scale-mzsi82utt#faq","mainEntity":[{"@type":"Question","name":"How many sandboxes does DeepSeek's agent training platform run at once?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek's platform, described in its DeepSeek Elastic Compute paper posted on arXiv on September 19, sustains about 380,000 concurrent sandboxes, with one production unit handling roughly 3 million sandboxes a day and more than 5,000 sandbox creations per second. About 130 people, including founder Liang Wenfeng, are credited on the paper. Anyone building agent execution infrastructure can track scaling approaches like this one on daily.dev."}},{"@type":"Question","name":"What isolation levels does DeepSeek use to contain misbehaving AI agents?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek's platform offers four tiers of isolation, ranging from function-call-level sandboxing up to full virtual machines. The paper states that agent execution is untrustworthy since agents may corrupt filesystems, exhaust resources, or interfere with system components, and argues no single mechanism can prevent all agent misbehavior, so the team instead strengthens observability and hardens the platform as models change. Teams designing containment for autonomous agents can follow approaches like this on daily.dev."}},{"@type":"Question","name":"When must EU member states have an AI regulatory sandbox operational under the AI Act?","acceptedAnswer":{"@type":"Answer","text":"Under Article 57 of the EU AI Act, every member state must have at least one regulatory sandbox operational by August 2, 2027, allowing developers to test innovative AI systems under supervision from national authorities. Member states may also choose to run a sandbox jointly with other states rather than independently. Developers navigating EU AI Act compliance deadlines can stay on top of rules like this via daily.dev."}}]} ```