Loading video...
Video Failed to Load
๐ Today we're launching Agent Sandboxes in Pydantic Logfire: a profoundly different kind of sandbox. "The world doesn't need another sandbox" they say. Maybe. Or maybe the world doesn't need another firecracker/gVisor/VM infra-fest. Our Agent Sandboxes are different. They're just an isolated interpreter with cold start times measured in... show more
1,514,946 views โข 2 days ago โขvia X (Twitter)
44 Comments
It's powered by Pydantic Monty, our open-source interpreter written in Rust, already used by @fastmcp, @AWS Stands, @Microsoft Agent Framework, @huggingface's MCP server, and and many more.
@pydantic Very cool, congrats! Did you actually drive the excavator ๐
@pydantic No, it's AI sadly.
@pydantic crazy video, love it
@mmatthias @pydantic Congrats Samuel!
@pydantic saweeeeet - excited to try it out
@zeeg @pydantic CoreWeave Sandboxes welcomes you to this space!
@pydantic congrats!
@pydantic i wanted to try this for a codemode use-case, but then it couldnโt run because we have a Rust dependency in our Python SDK ๐คทโโ๏ธ
@pydantic Just register an external function that calls the rust code.
@pydantic where does the external function run then?
@pydantic Code mode gonna raise further up
@intellectronica @pydantic Love the video!
@pydantic Resuming a REPL weeks later for virtually nothing is such a sharp feature. ๐
@pydantic I was always curious what would be the the ideal use case for Monty. Time to experiment with the new pydantic ai toolbox.
@pydantic this is incredible @pydantic !!!
@pydantic Congrats on the launch. Sandboxes limit what an agent can break; the other half is what it can reach. Watch for the "lethal trifecta": untrusted input, access to sensitive data, and a way to send it out. Take away any one leg and that attack path closes. ๐
@pydantic awesome launch video
@pydantic lol great video, congrats on the launch!
@pydantic @bot we need to test this! Find some time on my calendar so we can thoroughly test this brilliant creation @samuelcolvin and @pydantic
@pydantic @bot thanks!
@pydantic @bot Thanks to you. This is going to be awesome ๐
@pydantic This sounds rather amazing, will definitely give it a shot.
@pydantic I love it... ๐๐๐
@pydantic Is this Monty?
@pydantic Yes.
@pydantic Nice!
@pydantic excited to give it a shot!!
@pydantic awesome!!
@pydantic Well I should join Pydantic at this point!
@pydantic @samuelcolvin great to see monty graduated ! we're very happily using it in
@pydantic Login with gmail doesn't seem to work.
@pydantic ้ๅ ๅจ่ฟ่ฝๆฅ็ๅๆฅ็่งฃ้ๅจ็ถๆ็ปง็ปญๅ๏ผ่ฟไธช็จ้ๅพๅฎๅจใ้ฟไปปๅกไธญๆญๅ๏ผไธ็จๅ ่ฑๅๅคฉๆๅทฅไฝ็ถๆ้ๆฐๆญ่ตทๆฅใ
@pydantic Interesting, how much is this better than v8 isolates ?
LLMs generally find Python easier to write than js/ts (except for UI rendering). Memory usage should be lower. v8 doesn't support full suspend and resuming with serializable snapshots. v8 is ultimately 400k lines of c++ in the chromium codebase. Monty is ultimately ~100k lines of purpose built modern Rust.
@pydantic Hmm, reusable snapshots !
@pydantic Trying them now, I will compare it with E2B and Smol with MicroVM !
@pydantic the snapshot part is the one i'd use. when a repl resumes weeks later, what happens to the tool credentials it had, does it re-auth or fail closed?
@pydantic How would you compare this to OMP
@pydantic Pretty different. Agent Sandboxes isn't supposed to give a full development environment for the agent. Just a place to run python code, read/write files and call some external functions.
@pydantic Would OMP in a dicker approximate this?
@pydantic *docker
@pydantic Insane Launch Video ๐ฝ๏ธ
@pydantic Least privilege inside the sandbox is the right call, the gap opens when the agent steps out of it to move money. A session key scoped to the same allowlist would let it pay only for the tools it was given.
