Distinguished engineer at AWS. AI, agents, databases, and serverless. Views are my own.

Firecracker is intentionally feature-light. There are a lot of alternatives with richer feature sets (notably QEMU), many of which predate Firecracker. But small has its advantages, both in security and density.
i still do not understand why so many sandboxes use firecracker apart from being trendy, it lacks clear features like vfio-pci passthrough, virtio-fs or live migrations AI made it possible for a lot of people to roll infrastructure. But good infra will still differentiate.
10
6
325
22,025
I suspect the ideal would be a very pluggable microkernel-like core and then just the right features for the job at hand. Doing that without compromising performance or security is a little tricky, but certainly possible.
1
28
1,802
The Strands Harness team wrote a deeper dive on how we compared our harness to alternatives, running over 94,000 trials, achieving 34% lower cost on average. Don't miss it: strandsagents.com/blog/we-ra…
3
10
83
5,803
Looking for a great new decider model? Today, we're pleased to share four new strands-decider variants: more accurate, and better calibrated.
6
11
116
14,682
Check them out on HuggingFace: huggingface.co/StrandsAgents And GitHub: github.com/strands-labs/stra…
2
1
6
1,646
I love virtualization. You can see my name on the Firecracker paper. I've written a bare metal VMM using the CPU manuals (not a very good one, but it works). We picked it for AgentCore Runtime. So why didn't we pick virtualization for Strands Box?
Replying to @MarcJBrooker
curious why strands didnt use a cross-platform vmm?
7
22
419
31,038
(Common-sense policy? Starting to sound like I'm running for office here.)
1
21
2,770
We're not done with the whole vision of Strands Box by a long shot, but a lot of the future work comes down to policy. To controlling what agents can do in a powerful, deterministic, way.
1
1
20
2,673
Marc Brooker retweeted
Another day, another release from the Strands team! Strands Box was just released as a developer preview. Most sandboxes are just about what files can be read and if there is network access. This isn't enough granularity for many uses though. Strands Box uses kernel isolation to make a user-space policy layer the only way out of the isolation. This is enforced with Dogwood rules for shell use, Python, MCP and network access. One example shown is for git push being blocked unless npm test recently passed. Check this out from @MarcJBrooker. If you haven't already checked out Strands, give it a try. lckhd.eu/DXfzKZ #Agents #AgentSecurity #Sandboxing #StrandsAgents #AWS
3
6
25
3,362
A few ways: 1/ What it's for. AgentCore Runtime (the 'sandbox') is for running agents in the cloud, providing scalable, secure, highly available compute infrastructure. Strands Box is a local sandbox, running on whatever compute you already have (e.g. your laptop).
Replying to @MarcJBrooker
how is this different from agent core sandbox?
1
4
71
7,642
5/ Size of the TCB. AgentCore Runtime's is quite small (Firecracker, KVM). Strands Box's is quite big (host kernel, Python and Shell runtimes, Box's policy layer, etc).
1
2
689
You can use them together, running Strands Box inside AgentCore runtime, if you want. That allows both local and remote control.
1
4
647
This is one of the core reasons we released a developer preview: for complex local tools like git, command line filtering is hard to get right. A richer model of what git is doing makes writing simple correct policy easier, and we're working on just that.
Replying to @MarcJBrooker
the git push example also lets git commit -a through: it stages edits made after npm test without a git add, so nothing resets the window. keying it on any file write under the project since the passing test would close that
4
23
3,608