Environments · Terminal-Bench on Harbor

Import any Docker image. Zero broken environments.

Terminal-Bench is a widely used benchmark that drops an AI agent into a terminal and scores it on real tasks. Every task ships as a Docker image that defines its exact environment, and that is where sandbox platforms usually stumble: they translate Docker images into microVM images, and sometimes the translation quietly breaks things. A missing file here, a wrong permission there, and the agent fails through no fault of its own. The Docker image import we shipped last month does a faithful conversion that preserves the image's files and settings, so a sandbox behaves exactly like the container the task was published with.

To prove it, we ran Harbor, the framework behind Terminal-Bench, with its oracle agent across all 89 tasks. The oracle replays known-correct solutions, so it is not testing a model; it is testing the environment. Any task it fails means the environment is broken. On Tensorlake it passed 89/89 in 36 minutes 44 seconds, about 1.7× faster than Modal, which broke three environments. Pre-imported images boot in seconds instead of being rebuilt for every run.

Read the full write-up on the blog.

Engineering · Sandbox networking

We rebuilt how traffic gets into sandboxes. Throughput more than doubled.

Sandboxes talk to the outside world constantly: you upload files for an agent to work on, download what it produces, and call the APIs it serves. Every one of those requests used to pass through two proxies on its way in: an edge gateway that terminates TLS (Transport Layer Security, the encryption protocol behind HTTPS) and authenticates the request, then a second reverse proxy on the machine running your sandbox. Each hop parsed the stream and buffered it in userspace, which is pure overhead, and on large transfers it was the bottleneck.

The rebuild follows one rule: terminate TLS once at the edge for authentication, then move the bytes through the kernel. The second proxy is now a thin forwarder that never reads your data. With kernel TLS (kTLS), the kernel owns the TLS record layer, so encryption and decryption happen right on the socket. And splice(2) moves bytes straight between sockets without ever staging them in userspace. On 4 GB transfers, throughput went from 1.12 to 2.50 GB/s at roughly half the CPU per gigabyte. A side benefit: the forwarder now runs as its own daemon, separate from the process that orchestrates sandboxes, so routine dataplane deploys no longer interrupt live connections.

The details are in Zero-copy TLS ingress with kTLS and splice(2) for sandboxes.

Integrations · Devin Outposts

Devin plans in Cognition's cloud. It works in yours.

Devin is Cognition's AI software engineer: you hand it a task and it plans, writes code, and opens the pull request. Outposts is Cognition's framework for letting Devin work outside its own cloud, and we built an open-source reference implementation on Tensorlake. Devin keeps doing its thinking in Cognition's cloud, but every command it runs and every file it edits happens in Tensorlake sandboxes you control. Each session gets its own sandbox, booted with your private certificate authority (CA) certificates, your preinstalled tools, and your repositories already cloned, with access to internal services a hosted runtime can't reach. And when a Devin session sleeps, its sandbox suspends with memory, disk, and running processes preserved, then resumes from that exact state. Idle sessions stop burning compute.

Clone devin-outposts-tensorlake to run it today, or read the announcement post and the integration docs.

Integrations · OpenCode

Run OpenCode with its hands in a sandbox.

OpenCode is an open-source coding agent that lives in your terminal. Normally it runs commands and edits files directly on your machine, which is exactly as risky as it sounds. The tensorlake-opencode plugin moves those hands into a Tensorlake sandbox: the model's reasoning stays local, but every shell command, write, and edit executes in a disposable remote environment. A bad command never touches your laptop. Named sandboxes persist across OpenCode restarts, so installed dependencies, caches, and workspace files are still there tomorrow. Setup is one line of config and an API key.

See the OpenCode integration guide.

Try BYOC

We're offering free credits to developers who try Tensorlake and share their feedback. Whether you're running sandboxes in your own cloud with BYOC, pointing Devin at your infrastructure, or scaling parallel workloads, we want to hear what you build and what we can improve.