Applied AI @Tensorlake | agents, sandboxes, orchestration, multimodal LLMs
Dallas, TX
Joined April 2014
- Tweets340
- Following1K
- Followers306
- Likes1.4K
Shanshan Wang retweeted
You can now vertically scale running sandboxes on @tensorlake. This was one of the most requested features for building coding agents. Sandboxes can now adapt to the workload on the fly, instead of relying on static resource configuration.
The live DuckDB workload in the demo scaled up by 7x in throughput when its CPU and memory were increased live. If the machine running the sandbox doesn’t have enough capacity, the sandbox is transparently migrated to a node that can meet the new resource requirements.
Added @tensorlake as a sandbox provider to @scale_AI AgentEnv Framework. I tested it with a PortSim task:
An agent runs a port through a real windy week inside two Tensorlake sandboxes. Got a perfect score.
It was super fun to build and a great environment to play with.
The environment is the plugin github.com/earakely-scale/ag…, built on PortSimEnv by @adithya_s_k
Shanshan Wang retweeted
When building RL tasks and evals, it's difficult to distinguish between actual model capability failures and bad RL task design.
We're sharing some our task design principles so you can write better tasks before ever evaluating a model on them.
hud.ai/research/what-is-a-go…
Shanshan Wang retweeted
We are rolling out our new VMM to a small percentage of machines in production today!
It brings p95 resume latency for large sandboxes down to 1.5–2s when the snapshot isn’t cached on the machine.
In addition, it supports live vertical scaling of CPU, RAM, and even disk!
We’re wiring in live migration as well, enabling vertical scaling even when the machine running a sandbox doesn’t have enough resources.
We worked extremely hard on this over the last few months, and we’d started to hit the limits of what our previous VMM infrastructure would let us build.
Shanshan Wang retweeted
We are working on supporting the Agent Sandbox Protocol (ASP) in Harbor!
We have an active RFC for an Agent Sandbox Protocol (ASP) that disentangles an agent's execution environment from its runtime. There are many advantages it provides, including preventing trajectory modification.
- increased security
- resource isolation
- no contamination
- separate configuration (e.g. no network in execution environment)
For this to work, we'd need agent developers to opt in and implement their tools in a sandbox-agnostic manner using the protocol. I personally think this is the direction the industry needs to move in, and would love the community's help in designing this protocol.
github.com/harbor-framework/…
Got @tensorlake sandboxes working as ASP (Agent Sandbox Protocol) sandboxes in Harbor, using the SSH design. The agent runs on my laptop, and its tools run in a Tensorlake sandbox with the internet fully off. Tensorlake's native sandbox APIs can support other ASP designs too.
We have an active RFC for an Agent Sandbox Protocol (ASP) that disentangles an agent's execution environment from its runtime. There are many advantages it provides, including preventing trajectory modification.
- increased security
- resource isolation
- no contamination
- separate configuration (e.g. no network in execution environment)
For this to work, we'd need agent developers to opt in and implement their tools in a sandbox-agnostic manner using the protocol. I personally think this is the direction the industry needs to move in, and would love the community's help in designing this protocol.
github.com/harbor-framework/…
ASP proposal in Harbor by @kobe0938: github.com/harbor-framework/…
Shanshan Wang retweeted
@computesdk Actions 🤝 @tensorlake
Now you can run your actions directly on Tensorlake Sandboxes.
Tensorlake has worked tirelessly to provider elastic compute. Your CI/CD gets all of the benefits of their work now.
Shanshan Wang retweeted
Post-train an open model into a coding agent that accelerates developer workflows on everyday hardware.
The Gemma 4 Developer Agent Competition, hosted by @googlegemma and Kaggle, is live!
💰 Total Prize Pool: $100,000
⏰ Entry Deadline: November 25, 2026
Used Jev as a rule grader on this dataset. Basically I turned the labeling rules into yes/no questions and had it check the training data.
It agreed with hand corrected labels ~97% of the time, and actually caught a few inconsistencies in our own labels.
It doesn’t know the domain conventions. I put the convention in the question, and it applies the rule to the text.
A pretty lightweight way to audit if the labeling process is actually consistent.
Ran the best frontier models on industrial inspection notes and pick labels. Blindly but with full labels list. It’s already like an open book test but they still missed a lot.
Reading and understanding the input text is not a problem, but the right label depends on conventions only the domain experts know. And even worse, the conventions aren’t always clear or consistently documented.
So It’s always how fast you can consolidate and inject this knowledge, still a data problem at the end.
Got streaming working with @tensorlake sandboxes on Harbor too. So you can see the agent's trajectory and files while it runs. PR is on the way.
One of the first features we built into Terminal-Bench (long before Harbor existed) was --livestream to watch the tmux session.
Glad to see @kobe0938 cooking up the better, more generalized version for Harbor :)
Shanshan Wang retweeted
In the latest @harborframework release, dynamic network policies now work with Tensorlake sandboxes.
Harbor lets an eval change network access across different phases of a trial:
setup → public network
agent → allowlist
verifier → no network
We updated the Tensorlake Harbor implementation to support this directly, so the same sandbox can move through each phase with a different network boundary without restarting or losing state.
More on how it works 👇
Shanshan Wang retweeted
The full breakdown covers the mechanisms, trade-offs, and 4 questions to ask about any sandbox ingress hop.
Save or read on @towards_ai:
yadavdivy296.medium.com/ai-s…
Shanshan Wang retweeted
Worked a ton the last few days, but still feeling I have a lot of energy. Tensorlake had the most traffic this week than ever!
Our platform is not yet running at the scale of the schedulers I built before in my career at Netflix, Hashicorp or Facebook but we are getting there :)
Shanshan Wang retweeted
@cursor_ai Cloud Agents now run on @tensorlake sandboxes.
Start 1 worker or fan out many. Each gets its own fast, isolated sandbox.
Enterprise teams can also run Tensorlake in their own VPC, so agents can reach private repos, APIs & internal services.
Shanshan Wang retweeted
Tensorlake is back at #1 on ComputeSDK storage. 🥇
Storage lifecycle → 140 ms download, 598 Mbps
Snapshot & fork → 40 ms snapshot, 210 ms fork
Agents constantly save state, branch & run in parallel.
Fast storage primitives matter.
