Founder @tensorlake, building sandbox infra for agents. Past - Built Nomad @hashicorp, AI Infra @meta, Container Scheduler tech lead @linkedin and @netlfix
San Francisco, CA
Joined December 2007
- Tweets9.4K
- Following939
- Followers5.4K
- Likes4.4K
This is a big deal if they offer efficient snapshotting and live migration primitives. It’s hard to make the unit economics of sandboxes work on AWS without those.
it can not be overstated how big of a deal this is.
aws.amazon.com/about-aws/wha…
If @AnthropicAI gave everyone access to Mythos, the good guys would have had some weapons also in the cyber war.
We had a security incident on Tensorlake today. There was no impact to customer information such as API keys or sandboxes.
An application running on an upstream provider spins up sandboxes in our engineering team’s accounts and runs automation workflows. These sandboxes had access to our GitHub repositories.
The attacker pushed malicious code to branches and triggered GitHub Actions to probe for secrets that workflows can access to build and deploy code.
The upstream provider had a security breach a couple of weeks ago that exposed our environment variables. We had not rotated those env vars, as this is not a critical application for our production environment. This was an oversight, we should have acted on this immediately.
We have no evidence that any critical systems were accessed beyond probing for secrets. We have rotated all secrets and undeployed the internal application from the upstream provider.
Out of abundant caution, please rotate your Tensorlake API keys.
We will share more information with customers over email and post updates in our trust center. We have already notified all paying customers by email, and we will follow up with additional details.
Wow @coreauto keeps on hiring the best talent!
My dear friends, I’ve got 3 pieces of news
First off, I’ve retired from @midjourney after 3.5 years. It’s been an honor; we've pulled off great projects, and a few open-sourced pieces like Pretext & Handwired have been well-received. I’d do it again with the team, but they’re in a great stable place now and I planned to take >6 months off
Second, unfortunately the words spread. Life comes at you fast and (this is a crazy story best told in private) I was taken hostage by the good folks at @coreauto. I started last week. Retirement will have to wait for another few years
Third, I’ve figured out a few pieces of technology I wanted to exist for a while now. Some more like Pretext, some more like Handwired, and a mysterious new thing (pic attached. First one to guess right gets a repost). Stay tuned as we start training these at Core Auto.
We’ll show the world what AI can truly achieve in the hands of kind, capable & grounded folks who choose to look up.
P.S. we’re hiring one (1) backend engineer. Only requirement is showing us your previous project’s dependencies list
9
1,072
Part of building Tensorlake the past few months has been working with users at the frontier and the founders we get to partner with!
Our small team works extremely hard to keep on innovating and pushing reliability and what people can do with sandboxes every week.
Very cool product and powered by @tensorlake!
Today we’re introducing Aether @aetherruntime.
The fastest cloud coding agents in the world that you can use with your existing subscription
Give it a task and just 20 seconds later your build is running on its own cloud machine. You can pick up from your phone, watch every line, and step in whenever you want as if it was local. Close the laptop. It keeps going.
It runs Claude Code or Codex on the Claude or ChatGPT plan you already pay for so there’s not a 2nd AI bill. You only pay for the machine.
When it’s done, if code review is enabled, a reviewer agent runs the code with the changes, which catches classes of bugs that can't be found from static analysis.
Before building Aether I built infra at Apple, Uber, and HeyGen. I built Aether because the cloud agents I tried before all had drawbacks, and I couldn't imagine building something without it now.
It’s free to start. runaether.dev
There’s been a lot of chatter about CPU work stealing in sandboxes, and it’s hurting CI workloads. I think we have an elegant way to address this on @tensorlake.
We’re going to experiment with two tiers of sandboxes:
- A new tier that charges for requested CPU, memory, and disk, and does not over-subscribe any resources on the underlying machine. It will be more expensive, but premium with zero resource contention inside the sandbox.
- A second tier where we do over-subscribe resources to pack sandboxes densely, while continuing to offer active CPU usage based pricing.
Live migration will allow users to move between tiers transparently while a sandbox is running.
Truly don't understand how cheap sandboxes work when performance of blob store (and colocation with compute), bandwidth, SSD, etc, costs a bunch of money.
There are now 10 sandbox companies claiming to be the cheapest.
The cheapest: @Hetzner_Online
I was doing some research this evening on how swarms of agents can share lessons among themselves and improve their effectiveness over time, both in implementing features and operating software in production.
I’m specifically interested in whether we can incorporate an understanding of real-world performance (once software is deployed) into the sessions where code is written.
Some interesting papers and blogs in this area:
If Apple one day lets companies to run as many VMs they can in a single host, I don't think MacBooks as a product line will exist. iPad + MacOS VMs are good enough for most use cases that people have for MacBooks. I think the limitations around number of VMs per host is intentional to keep the product line going.
One of the tweaks we recently made in our internal software factory, Foundry, is giving the agents the ability to spin off new sessions by cloning the sandbox it’s running in. This lets agents run many investigations and implementations in parallel.
We’ve also exposed tools that allow sessions to discover other sessions and send messages back and forth, creating a community of agents learning from each other and sharing outcomes from performance benchmarks, etc.
Turns out agents LOVE cloning themselves and messaging each other to coordinate work! They seem to be good at referencing work done by other agents in the community and also send each other relevant tasks related to the area of code they were working on.
With copy on write/forkable compute architecture, I think staging becomes ephemeral - clone a whole environment, run tests and throw it away. Every developer/agent gets it's own environment to test new algorithms and features.
One of the coolest thing about building sandbox infrastructure is how fast customers scale once they like the platform!
@TaskletAI is already at 50k+ sandboxes on @tensorlake and that's about half of their workload!
(screenshot shared with permission)
This is great! Some of my thoughts on this -
1. The idea of "staging" is still going to be there, but coding agents would be able to spin up VPCs of services in sandboxes in seconds to test new changes.
2. Environments like prod/dev would be forkable - I think sooner or later sandbox companies will offer this. We are already working on it.
3. The DST market will grow - she is 100% right about this. When agents work on mission critical systems like databases and file systems, micro optimizing on one use case often leads to regressions in other areas. Similar thing with state machine based systems, easy to miss what an error could do to the state transition of the system.
My further (informed) fantasies about the near future:
1) No CI. Agents prove each change against rules while they write it.
2) Code is a cache: rules stay, code regenerated with better models.
3) No staging, simulation instead.
4) Humans direct in plain words, agents compile that into formal rules, humans don't read the formulas, only examples and counterexamples.
5) Agents find the rules the human never knew to express (and ask to confirm them).
6) The checker catches contradictions between human's rules while they're still directing.
7) Unchecked/unproved code becomes legacy code.
8) Branches come back as experiments (many versions at once, humans pick among winners).
9) Software gets bred, not built.
10) Slop is variation without selection.
11) If UI is generated on the fly, design becomes the rules every generated screen must follow. A checker runs those rules on every generation.
12) Apps evolve from how they're used.
13) Everyone uses their own software version. One product becomes many personal versions, that stay compatible via shared rules.
10
19
1,814
This event is tonight at @recallai's office! If you are into sandboxes come and say hi.
Giving a talk at Unfortunately Computers on October 29th about @tensorlake’s content-addressed block storage system and efficient snapshotting infrastructure for sandboxes.
If you love sandbox tech, please come say hi! unfortunatelycomputers.com/e…
Thanks @davidgu and @recallai for hosting.
1
9
1,086
This is cool!
We just open sourced our internal design brain 🧠
Meet Content Studio - the agent behind 3,000+ assets at @LangChain
Create on-brand product launch videos, slide decks, ad creatives, social media assets, ui mockups and more
Storyboard and collaborate on a shared canvas in @paper and edit videos in code with @HyperFrames_
It even made its own launch film below, and went full Rocky 🥊
Fork it, add your brand guidelines, and deploy it to your team in Slack with Managed Deep Agents
Here’s what 's inside 🧵
This makes no sense. We offer active cpu usage based pricing. Daytona and E2B can’t be cheaper than us 😂
Having said that, the goal is to make the most premium product not the cheapest.
had agents running last friday in a bunch of sandboxes
i forgot to turn them off during the weekend, but it was so cheap on boat.dev that i didn't care
if you're not using boat you are literally lighting money on fire for zero reason
So basically everyone is converging on - sandboxes, CI, apps, git, versioned file systems and managed inference?
Introducing Vercel CI
• Intelligently skips tasks with Turborepo
• 2-6× faster CI in our internal repositories
• Optimized for agents via Vercel MCP and CLI
After nearly 8 million tasks, we're excited for the future of agentic CI. In beta on all plans.
vercel.com/blog/introducing-…
