[KongchangAI]
Product

Metal Sandbox

An isolated execution and scheduling infrastructure for GPU coding agents on Apple chips. It provides filesystem-level workspace isolation, native command execution, and command-level scheduling across a cluster of Macs.

Timeline (last 90 days)

Nov 13

Metal Sandbox is an infrastructure for GPU coding agents that integrates per-instance file access restrictions, native command execution, and command-level scheduling across interconnected Macs.

Unverified50%
Nov 13

Metal Sandbox's isolation strategy is built at the file system level: each agent instance can only access authorized files, and its development environment is also private.

Unverified50%
Nov 13

In Metal Sandbox, parallel commands can run simultaneously, while serial commands share a single execution slot on each node.

Unverified50%
Nov 13

Metal Sandbox supports command migration: if a command queued on a node meets the migration conditions, it can be scheduled to execute on an idle Mac, with the result returned to the original agent.

Unverified50%
Nov 13

Metal Sandbox integrates with Harbor, supporting task execution, automatic verification, and rollout collection through a unified workflow.

Unverified50%
Nov 13

Controlled experiments on four Mac mini M4 nodes showed that command throughput improved by 1.68x to 2.05x under centralized demand.

Unverified50%
Nov 13

Estimates based on sixteen static agent trajectories indicate that, assuming sessions are continuously replenished and average duration stays unchanged, four sessions per node can achieve about 2.14x sustained throughput gain compared with a single session.

Unverified50%
Nov 13

Trajectory replay under dynamic scheduling showed that average tool wait time dropped by 18.8% and batch duration decreased by 22.3% (assuming stage durations stay unchanged).

Unverified50%
Nov 13

Metal Sandbox and its Harbor integration have been open-sourced on GitHub.

Unverified50%
Nov 13

What Metal Sandbox addresses is not a model capability problem but an engineering efficiency problem.

Unverified50%

All Facts (10)

Unverified

What Metal Sandbox addresses is not a model capability problem but an engineering efficiency problem.

50%
Unverified

Metal Sandbox's isolation strategy is built at the file system level: each agent instance can only access authorized files, and its development environment is also private.

50%
Unverified

In Metal Sandbox, parallel commands can run simultaneously, while serial commands share a single execution slot on each node.

50%
Unverified

Metal Sandbox supports command migration: if a command queued on a node meets the migration conditions, it can be scheduled to execute on an idle Mac, with the result returned to the original agent.

50%
Unverified

Metal Sandbox integrates with Harbor, supporting task execution, automatic verification, and rollout collection through a unified workflow.

50%
Unverified

Controlled experiments on four Mac mini M4 nodes showed that command throughput improved by 1.68x to 2.05x under centralized demand.

50%
Unverified

Estimates based on sixteen static agent trajectories indicate that, assuming sessions are continuously replenished and average duration stays unchanged, four sessions per node can achieve about 2.14x sustained throughput gain compared with a single session.

50%
Unverified

Trajectory replay under dynamic scheduling showed that average tool wait time dropped by 18.8% and batch duration decreased by 22.3% (assuming stage durations stay unchanged).

50%
Unverified

Metal Sandbox is an infrastructure for GPU coding agents that integrates per-instance file access restrictions, native command execution, and command-level scheduling across interconnected Macs.

50%
Unverified

Metal Sandbox and its Harbor integration have been open-sourced on GitHub.

50%

Source Articles