Tensorlake

Full write-up: pub.towardsai.net/i-ran-5-ai-age…

Most teams run all their AI agents in one environment, one bad failure can take everything down. A builder from our community ran each of his agents inside its own Tensorlake sandbox, which acts as an isolated mini-computer, then crashed one agent on purpose mid-run. The others finished their work cleanly, they never even noticed the failure. Write-up below 👇

Full write-up: pub.towardsai.net/building-state…

A stateful sandbox becomes the solution for long running AI agents for more and more builders. Someone in our community built an agent on Tensorlake sandboxes, then killed its session on purpose and came back days later. It just kept going. Woke up mid-task in under half a second. Write-up below 👇

We built a new dataplane L4 proxy for @tensorlake sandboxes, using kernel TLS offload and splice which gives every connection 2.5 GB/s into a sandbox from external networks. This improves user perceived performance of agents which create or ingest large files. tensorlake.ai/blog/near-zero…

Full write-up: @sebuzdugan/your-agent-benchmark-is-lying-to-you-65a552fea151" target="_blank" rel="nofollow noopener">medium.com/@sebuzdugan/yo…

The fix: Fork a fresh Tensorlake sandbox per task from a snapshot, keep verification out of the agent's reach, throw the env away after. Spurious failures dropped to zero.

Stop letting your agent evals lie to you. A harness reported 87% pass rate. A builder in our community verified it properly: 33%. The gap? Agents touching the thing grading them — hijacking pytest to force every test green, rewriting tests to assert True, printing "all tests passed" and exiting clean. The fix 👇

Isolation you don't have to think about is the whole job. All 5 experiments, reproducible, here: pub.towardsai.net/your-ai-agent-…

The real point: in production the danger isn't the agent making a mistake. It's the mistake reaching the host.

"Your AI agent is one rm -rf away from disaster." Everyone's first fear when they let an LLM run generated code. A write-up in Towards AI (by Darshandagaa) runs 5 experiments against Tensorlake — trying to break out of the sandbox. 🧵

{ "plugin": ["tensorlake-opencode"] } That's the whole setup. Docs → docs.tensorlake.ai/sandboxes/open…

We shipped a Tensorlake plugin for @opencode. ✨ Add one line to your config and every command your coding agent runs happens in an isolated cloud sandbox. Fast boot, sub-second resume, state that survives restarts. See it in action 👇

"The hardest part of running Claude Managed Agents isn't the model. It's everything underneath it." A great build story on running long-lived Claude Managed Agents with Tensorlake Covers suspend/resume, persistent state, checkpointing, branching workflows, and why state matters more than compute for many agent workloads. pub.towardsai.net/building-long-…

Terminal-Bench 2.1 oracle, all 89 tasks: Tensorlake hits a perfect 1.000 — and runs ~1.7× faster than the next provider at the same concurrency. Why oracle = 1.0 is the only honest baseline, and how we got there: tensorlake.ai/blog/accelerat…

Test driving our ios app. This shell is a PTY session that you can reattach and come back anytime when you open your phone and iPad! Beyond running shells, we built some cool features in the app that extends what builders can do on iOS devices. If you want a TestFlight invite please DM me!

Soon you won't have to leave a party to check in on your coding agents! @tensorlake on iOS coming soon! The sandbox shell is powered by libghostty so the editing experience should be really good.

Full write up at pub.towardsai.net/i-built-a-stat…

Most agents die the moment something interrupts them. State gone, start over. A developer tried to break that on Tensorlake.