Blaxel

We just shipped the product we're most excited about since we launched Blaxel: Agent Drive. tl;dr: It’s a shared filesystem for AI agents. Mount it to multiple sandboxes at the same time, with concurrent read-write access. Write a file in one sandbox, read it from another. For agents, it becomes a shared memory layer that lives outside any individual sandbox. https://lnkd.in/eAUTf7Ms Here’s why we built it. Over the past year, we’ve worked with hundreds of teams running agents in production. The pattern keeps showing up. Once the harness and sandbox are solved, the next bottleneck is sharing context across runs. Teams end up building custom S3 or GCS export-and-import pipelines. Or they accept that every agent session starts from scratch. That means redundant token usage, slower resolution times, and engineering time spent on plumbing instead of product. Agent Drive removes that handoff layer. Agents share a filesystem the way teams share a Google Drive folder, except it is purpose-built for agentic workloads: - Small-to-medium files - High concurrency - Frequent reads and writes across multiple sandboxes Agent Drive is in private preview. If you are building multi-agent systems or agentic workflows that need shared context, I’d love to hear what your agents need to share. Docs: https://lnkd.in/eA_sa6d6

New: Blaxel Agent Drive. Our biggest product since we launched sandboxes. It is a shared filesystem for AI agents. Mount Agent Drive to multiple sandboxes at the same time. Write a file in one, read it instantly from another. Full concurrent access. Agent Drive is fully durable, and it scales elastically with no capacity limits. Most importantly, it is purpose-built for agentic workloads: small-to-medium files, high R/W concurrency. First beta-testers report 5-10x better performance vs. S3-backed alternatives. Available now in private preview. Join the waitlist in the comments.

Agent Drive is the product I’m most excited about since we launched Blaxel. Why? A sandbox gives an agent compute. But compute is only one part of the environment. When agents work across multiple sandboxes, they also need storage that persists beyond a run and can be accessed concurrently. This is Agent Drive: a shared filesystem you can mount to multiple sandboxes with concurrent read-write access. Write a file in one sandbox, read it from another. Without it, teams either start every session from scratch or build export and import pipelines around object storage. Agent Drive removes this handoff. It's in private preview. If your agents need to share context across sandboxes, I’d love to hear what they need to share. Demo: https://lnkd.in/eptqpR82 Docs: https://lnkd.in/eTTS8Aa6

CodSpeed built an agent that makes your code faster automatically with Claude Managed Agents + Blaxel (YC X25). Here's how it works! It ingests a customer's code, benchmarks it, iterates on changes, and surfaces real performance improvements. The agent harness runs on Claude Managed Agents. The execution layer runs on Blaxel sandboxes. The loop is simple: for each session, Claude Managed Agents spins up the agent, and Blaxel provisions a worker sandbox from CodSpeed's custom image. The agent works on the code in-sandbox until it finds improvements worth keeping. Results come back as a structured diff with benchmarks. Three things made this work: ➡️ Speed: Blaxel sandboxes suspend after 15s of inactivity and resume in 25ms, so CodSpeed waste zero time rebuilding repo context. ➡️ Security: the agent touches real, proprietary customer code, so confidentiality is non-negotiable. CodSpeed was already a Blaxel customer and trusted the infrastructure and the team. ➡️ Control: CodSpeed bakes its own profiling tools into the sandbox image, and configures its networking layer to inject secrets in outbound traffic on Blaxel. Control over the substrate was paramount. Check out our technical deep-dive with CodSpeed on the architecture in the comments!

Our YC batchmates at Casco (YC X25) just dropped one of the best technical write-ups I've read this year! René Brandel broke down exactly how North Korean operatives are sneaking malicious code into popular open-source libraries via legitimate-looking PRs. The attack is wild: obfuscated payloads hidden in build config files, blockchain used as permanent hosting that no one can take down, and zombie processes that keep running even after your build exits cleanly. 30+ repos already found with the same fingerprint. Probably hundreds more out there. If you ship anything that touches npm, this is worth 15 minutes of your time. (BYTW, the Casco team used Blaxel sandboxes to safely detonate and analyze the malware, which was cool to see in action 🧨.) Link in first comment 👇

Corvera (YC W26) is building the context layer for AI-native CPG brands, accelerating the adoption of agents that replace the manual layer around ERPs. A core product challenge: Corvera needed a fully isolated compute environment, spun up instantly with the entire workspace context, as soon as a customer initiated a conversation with their agent. With Blaxel, they got this and more: near-instant sandboxes for secure, isolated compute and a shared filesystem for persistent organizational memory. Today Corvera deploys ~1,000 sandboxes daily on Blaxel and has hit 2x faster performance than competitors across real-world agent workflows. A great example of how agent-native infrastructure unlocks agent-native products. Read the full case study in the comments.