01 agents
Coding agents
Give your agent a real machine — install packages, edit files, run tests, iterate. Isolated from your infra, destroyed when the loop ends.
- Full root shell
- Preinstalled agent images
- Persistent workspace
Solutions
Sandboxes built for
agents, LLMs & RL.
BoltzLabs is the execution layer AI systems need — isolated, disposable machines your code can create, drive, and destroy programmatically. No VMs to manage. No residue left behind.
lifecycle
01 create POST /sandboxes
02 run agent · llm · rl
03 destroy zero residue
02 llms
LLM code execution
Run model-generated code safely. Every tool call gets a fresh sandbox — no shared state, no escape to your host.
- Untrusted code isolation
- Language-agnostic runtimes
- Per-second after hour 1*
03 rl
RL rollouts
Spin up thousands of concurrent environments for training and eval. Ephemeral worlds that boot fast and scale with your GPU jobs.
- Massive parallelism
- Reproducible envs
- Batched step fan-out
04 async
Background agents
Long-running workers with the tools, context, and credentials already in place — a full dev environment your agent owns.
- Persistent workspaces
- API-driven lifecycle
- Reconnect anytime
05 evals
Evals & benchmarks
Identical environments for every trial. Fan out parallel runs, reset rollout workers, and score agents without flaky local setups.
- Warm resident envs
- Batched step fan-out
- Soft or hard reset
06 ci
CI & batch jobs
A fresh machine per job. Install, build, and test in isolation, then destroy the workspace when the job exits.
- One job → one box
- Any runtime image
- Destroy on exit
One API. Any workload that needs a machine. Create → run → destroy.
read the docs