Shanshan Wang

Building applied AI @Tensorlake | agents, sandboxes, orchestration, multimodal AI

[Diptanu Choudhury @diptanu]

Recruiting early design partners for new products we are building has been a pleasantly enjoyable job over the last 6-8 weeks :)

[Tensorlake @tensorlake]

"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.
https://t.co/gSQcjJvELo

[Diptanu Choudhury @diptanu]

We have been working hard on making BYOC a first class primitive on Tensorlake along side our SAAS offering.

We hear regularly from our users that they want to use our software stack on the compute that they buy from hyper-scalers and Neo Clouds. So we are making self hosted compute as first class as we can.

You can tie in a Tensorlake project to a self hosted cluster, and continue to use our cloud APIs with **ZERO** code changes. Sandboxes will magically appear in your infrastructure on AWS and GCP.

[Shanshan Wang @cooltxeel]

@tensorlake 📦 OpenIngest:
https://t.co/IC8CUqcck8

[Shanshan Wang @cooltxeel]

If you missed our June newsletter:
• We open sourced our Document AI product as OpenIngest:
• Going forward, the recommended way to run OpenIngest is on @tensorlake Compute.
• Sandboxes now support importing any Docker image (public or private).
• Added support for Claude Managed Agents.
Links in thread👇

[Shanshan Wang @cooltxeel]

@tensorlake 📖 June newsletter:
https://t.co/TTjmGYWhKg

[Shanshan Wang @cooltxeel]

We registered the Terminal Bench 2.1 task images in Tensorlake as public images. Anyone running Terminal Bench with @harborframework can use these images directly instead of rebuilding them from Dockerfiles.

[Tensorlake @tensorlake]

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:
https://t.co/QEhFF4VwrS

[Tensorlake @tensorlake]

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:
https://t.co/QEhFF4VwrS

[Diptanu Choudhury @diptanu]

There needs to be a sandbox benchmark which measures the interruptions to network connections and blips in sandbox availability for multi day agentic sessions.

Sandbox users will thank the creator of such a benchmark 😅

[ComputeSDK @computesdk]

🎉🎉🎉

Huge congrats to @tensorlake for successfully completing the 2026 100k Scale invitational!

See the full results below 🧵

[Diptanu Choudhury @diptanu]

The 100k benchmark numbers from @computesdk are in! I'll write a detailed post on how Tensorlake's architecture works across the control plane scheduler, the data plane that runs the VMs, and the network ingress that manages connections.

Bursting from 0 to 100k sandboxes without warm pools is a hard systems problem. The world deserves a good behind the scenes post :)

[Diptanu Choudhury @diptanu]

In case you missed it! We are shipping a new file system for GPU sandboxes which cold starts in 1-2 seconds even for CUDA and Torch images which are 10-15G in size.

[Tensorlake @tensorlake]

Awesome! Cool sandbox agent powered by @tensorlake and @AnthropicAI

[Shanshan Wang @cooltxeel]

Interesting observation:

Many multi-agent systems spend more time preparing to work than actually working.

Snapshots help because setup becomes a one-time cost instead of an every-agent cost.

Nice walkthrough of the pattern here.

[Divy @aiwithme0001]

Full breakdown (the architecture, the real code, and the 3 mistakes that cost me 3 rebuilds) is here:

https://t.co/Fh1dJWCLHQ

If you're building anything with multiple AI agents, check this before you touch anything else.

[Tensorlake @tensorlake]

Parallel agents don't automatically mean efficient agents.

If every worker rebuilds the same environment, you're still paying setup cost N times.

Snapshots + forks change the equation.

Interesting write-up on a pattern we're seeing more often in production agent systems.

[Tensorlake @tensorlake]

Importing public and private Docker Hub images into Tensorlake is now supported in both the Python and TypeScript SDKs, bringing the same functionality previously available in the CLI to programmatic workflows.
https://t.co/dDVf6iqZnz

[Shanshan Wang @cooltxeel]

The future bottleneck isn't execution. It's preserving and reusing work.
Nice to see people are building agents using stateful Tensorlake sandboxes

https://t.co/MFyLtIlxMn

[Diptanu Choudhury @diptanu]

@damianfu2 我们直接将现有的 OCI 镜像转换为 VM 镜像，无需您编写“模板”。此外，我们也支持基于 Dockerfile 创建 VM 镜像，这种方式更接近 E2B 模板。

https://t.co/3cy9jUmPkD

[Tensorlake @tensorlake]

OCI image support landed for Tensorlake Sandboxes in Harbor. A real eval task now goes 250s → 15s per run — you build once, then every rollout boots from the cached image. https://t.co/2aILtWEiMJ

[Tensorlake @tensorlake]

Just checked the latest @computesdk rankings before wrapping up the week.
@tensorlake ended with:
🥇 #1
🥈 #2
🥈 #2
Really proud of what the team has accomplished over the past few months. Onward 🚀