Developer kvcache-ai released AgentENV, an open source platform for running large numbers of isolated agent environments at scale, according to the project's GitHub repository.
The platform launches each environment in a Firecracker microVM, the same virtualization technology behind AWS Lambda, and loads container images on demand through overlaybd so it can serve more environments than local disk space would otherwise allow, the repository says.
AgentENV's environments boot or resume from a snapshot in under 50 milliseconds and pause in under 100 milliseconds, according to the repository, which lets idle environments release CPU and memory quickly instead of sitting reserved. The project reports native snapshot and fork operations that also complete in under 100 milliseconds, plus shared host page caching that sustains heavy memory overcommitment.
The repository says the platform was built to run agentic reinforcement learning training for Kimi K3. Its documentation notes that AgentENV "currently does not support authorization" and warns against exposing it publicly without additional security controls.
Reinforcement learning training and agent evaluation both live and die on how cheaply teams can spin up and tear down isolated environments. AgentENV's numbers target that cost directly, at the price of authorization being a problem builders must solve themselves before opening access beyond a trusted network.