Eran Sandler

Builder, operator and investor. Infra, AI, and product nerd. Trying to make powerful things simple. Opinions are my own.

Building

[Utpal Nadiger @utpalnadiger]

Timeout bound coding agent sandboxes make you babysit snapshots and preview URLs to keep an app "alive".

In https://t.co/Yz61qJLcD0, when idle, the box hibernates ($0 compute, URL intact).

Any request coming in auto wakes it.

Hibernation snapshots the VM state (memory + disk) automatically and stops the VM, so there are no compute costs while hibernated, you only pay for the stored snapshot.

Folks building lovable's for {x} love this!

https://t.co/bpPljACppq

[Eran Sandler @erans]

@jeresig @CloudflareDev not there yet, but works with Cloudflare R2: https://t.co/d8UT43KwL6 (expect full Cloudflare example soon)

I agree that we need more decentralization to deal with agents.

[Eran Sandler @erans]

@wernerk_au @trailofbits Not around skills specifically but something that can help with the aftermath

https://t.co/voUDo9sx6s and soon a local machine tool (type of EDR, but active and not passive) that can enforce it on all your local AI agents.

[Eran Sandler @erans]

With the abundance of open weights providers out there I really wanted to be able to have an API to compare the same model across providers as well as the ability to have all that information be consumed easily by LLM.

I've created https://t.co/FHKpevyRys for exactly that reason.

Any feedback is welcome!

[Trail of Bits @trailofbits]

We built four malicious skills to test whether skill scanners actually work. Three took less than an hour to conceive and implement. ClawHub, Cisco, and Vercel's https://t.co/nUlnRcQWyG marked them as safe. 🧵

[Eran Sandler @erans]

@bentlegen Single meta MCP server that you configure other MCP servers and you only need to configure it once!

[Eran Sandler @erans]

@mitsuhiko With AI it is now easier to find and fix these issues and send an upstream patch. I think the amount of work and dependencies made it hard before AI for smaller companies or products to do this.

[Eran Sandler @erans]

4/ This is what we’re building at Canyon Road: AgentSH at runtime, Beacon on endpoints, Watchtower across the fleet.

The agent proposes the action. Policy decides whether it happens.

https://t.co/bKJ5bDbdjg

[Eran Sandler @erans]

1/ The attack will not look like malware.

That’s the thing that keeps coming up for me with AI agents. The risky action may use approved tools, valid credentials, and workflows that look completely normal.

[Eran Sandler @erans]

3/ That’s why prompts, settings files, tool permissions, and logs after the fact are not enough. They matter, but the control point has to sit outside the agent’s reasoning loop, where the agent cannot bypass it.

[Jeffrey Reitman @JeffreyReitman]

1/
After founding a gaming company and spending more than a decade in venture, one thing has become very clear to me:

I love early-stage investing.

Today, I’m excited to share that I’ve started @entropyVC.

[Eran Sandler @erans]

@mlejva @e2b Persistence is key (literally now with AI agents compute :) - but also in keeping up with the vision)

[Eran Sandler @erans]

@rauchg AgentSH - mostly Claude (and some codex) https://t.co/voUDo9sx6s

[Eran Sandler @erans]

@samuelcolvin @JordiMonPMM Like anything new in the market, there is some level of education. AI is moving so fast that it's hard even for those of us in the weeds to follow it all.

While AI implications are not just technical, we are still at the point that understanding the tech and its piece matters.

[Utpal Nadiger @utpalnadiger]

@motatoeshq cooked with this. I've been hooked to clawputer over the last few days

https://t.co/9jKfXFZTro uses:

- https://t.co/qPwpgFRfs4 as the computer
- @pipedream for integrations
- @browser_use for browser use (lol)
- https://t.co/wnyKSYW4DI for imessage infra

[Eran Sandler @erans]

@utpalnadiger it removes the "stealing secrets" attack vector, but not the "use the secrets and that API to access the info, exfiltrate it or do some other harm. That is why agent actions needs to be managed, regardless of the prompt (injected or not) - why we built https://t.co/Zvm0T6rQt0

[Utpal Nadiger @utpalnadiger]

Firecracker doesn't support live migration.

It has snapshot/restore ie. pause the VM, save state, restore later. But it is not really the same thing.

True live migration means serializing a running VM (CPU, memory, devices) and moving it to another host mid execution. QEMU has done this for 10+ years w iterative pre/post-copy + dirty page tracking.

Firecracker was built for Lambda. Long running agents need both Lambda ergonomics WITH EC2 semantics. QEMU ftw.

[Eran Sandler @erans]

@mitsuhiko @bentlegen It's harder to keep a mental image of the code when you don't spend the time writing it, however, speccing it hard can help with that. When I look at the code I wrote 7 years ago I don't remember every detail but I can read it - or ask my clanker :)