Skip to main content
AgentEnv is an open-source framework from engineers at Scale AI. You use it to build RL environments, and to run and grade agents inside them. AgentEnv has a built-in boat sandbox provider. With it, each environment and each agent runs in its own boat sandbox: an Ubuntu VM with Docker inside.

​
Set it up

You need Python 3.11.4 or later and a boat account. For the trial, see Free trial.
1

Get a boat API key

Create a key on the API keys tab of the dashboard, or run boat api-key create agentenv --preset ci. Keep it in an environment variable:
2

Install AgentEnv with the boat extra

3

Configure the provider

Add the key reference to your AgentEnv config. Use env: or secret:, never the key itself:
.agentenv/config.toml
4

Check your setup

Run the built-in hello task on boat. It needs no model:

​
Use boat for your runs

Make boat the default for environments, agents or both:
.agentenv/config.toml
To choose boat for one run only, pass a flag: A comma-separated value, for example "boat,local", tries the next provider when one fails to create a sandbox.

​
What happens on boat

Each sandbox is a no-env sandbox. It gets none of your secrets, credentials or private repositories, and it cannot act on your account or on other sandboxes. Snapshots are off, so a stopped sandbox keeps no disk.

​
Sizes

AgentEnv asks for CPU, memory and disk. The provider picks the smallest boat size that covers all three: A request bigger than xlarge fails. See Machines.

​
Labels

Every sandbox carries the label agentenv.managed=true. AgentEnv attribution (for example [sandbox.attribution] in its config) becomes agentenv.attr.* labels. Use them to find a run’s sandboxes and to join sandbox usage to your runs.

​
Limits

When boat refuses a create because a start limit is full, the provider waits and tries again. It makes up to 6 attempts. Keep the number of envs and agents that start at the same time under your limits.

​
Not available

The provider refuses these with an error. It never creates a sandbox that is less isolated than you asked for.
  • An image override. Every sandbox boots the standard boat image. See Base image and tools.
  • A network_policy that restricts egress. boat has no egress allowlist yet.

​
Configuration

​
Resources

AgentEnv sandboxes

The boat provider in the AgentEnv docs.

Source on GitHub

The AgentEnv framework and its providers.

Evals and RL environments

Run many eval episodes on boat: saved snapshots, forks, limits and cost.

API keys

Create the key that the provider needs.