Sandbox primitives and systems-language control planes that own the agent loop

MaxGfeller

Code-first composable SDK built on Vercel AI SDK for powerful AI agents with sessions, middleware, VFS providers, and React/Vue hooks.

Use caseUse it to wire Agent plus createFsTools and createBashTool for sandboxed file and shell access inside your own product UI.

dhanababum

Easy microVM sandboxes for AI agents and untrusted code on AWS Lambda MicroVMs with Python SDK, asb CLI, and MCP server.

Use caseUse it to boot a Firecracker sandbox in seconds, exec agent code, forward ports, and suspend/resume long sessions without managing clusters.

vivek1504

Firecracker microVM sandboxing for AI agents with millisecond boots from snapshots, full network, and native MCP support.

Use caseUse it to give each agent session a dedicated hardware-isolated Linux machine that restores in under 100 ms end-to-end.

katosh

Kernel-enforced credential and network jail for untrusted Linux processes with Slurm awareness and built-in agent profiles.

Use caseUse it to wrap coding agents on HPC login nodes so sbatch and credentials stay sandboxed without containers.

Docker

September 24 2026 marks the experimental public release with a Kits catalog, Hermes and Antigravity curated kits, credential controls, custom secrets, and improved sandbox provisioning progress. Launcher now reuses saved keys and supports platform selection plus restart timers.

Shipped

Datadog

Engineering writeup on building agent harnesses around observability so traces, metrics, and feedback loops become first-class controls rather than afterthoughts. Argues that production reliability comes from instrumented runtimes that surface tool failures and context bloat early.

Read