Agent Rush

Checking the worker Inference not reported

Run AI-written code in a sealed box that is destroyed after every run.

An agent writes Rust for your task and runs it with no network. The box starts small, grows only as the code needs, and is checked clean afterwards.

Operator sign in

Only the operator can start tasks and read their receipts.

Safety and limits

Authority stays outside the agent. Only the trusted worker creates, resizes and destroys sandboxes; the agent can only ask.

Isolation
gVisor, a non-root user and a read-only system. No host folders and no Docker socket inside.
Network
Off. Cloud and inference keys never enter a sandbox.
Growth
Each run starts at 256 MiB and half a CPU. The worker raises them live when the code needs more, up to the ceiling the agent set for the task.
Hard limits
No ceiling can pass 1 GiB of memory, 1 CPU or 2 minutes per run. 64 processes, 64 MiB of scratch space and 64 KiB of output.
Files out
Regular files only: 1 MiB each, 4 MiB in total, eight at most.
Cleanup
Every sandbox is destroyed, and its memory stays reserved until the cleanup is confirmed.
Evidence
Records collected by the worker. They are not an independent attestation or a security audit.