Every RL environment we build runs on AgentEnv. Today we’re open-sourcing it, so more people can build the worlds agents learn in.

Learn more:

open sourcing the env layer is big

Open environments are where most teams can actually learn from agents. A good first step is rebuilding one real task from your own workflow inside it and checking where the agent's path diverges from how your team does it.

Big move. Open-sourcing AgentEnv will speed up how fast we get better RL worlds.

@Laz4rz you seen this?

Open-sourcing the env layer every RL environment runs on is a huge unlock. More people building the worlds agents learn in means way better agents, let's go!

Does AgentEnv cover non-RL agents too, or just RL?

impressive

cc @adithya_s_k

Open sourcing the environment layer matters more than most model launches. Loads of agent failures come from training worlds that look nothing like real, messy software. Keen to see envs for boring stuff like invoices and calandars, because that's where agents actually get used.

Nice detail in the docs: task definitions get a new version on every write, and each run gets its own record. Useful plumbing for asking whether the agent improved or the exam changed 😄

Building AgentEnv Framework with our team at Scale AI has been a labor of love. We will be sharing more soon; eager to see what y'all build with it!

nice. curious how you handle env state reset between rollouts, snapshot whole container or replay from a seed? that's usually where rl env infra gets slow

Nice detail in the docs: task definitions get a new version on every write, and each run gets its own record. Useful plumbing for asking whether the agent improved or the exam changed 😄

Per Scale's docs: Apache-2.0 license, 4 sandbox options at launch (local Docker, Modal containers, Modal or E2B VMs), tools via MCP/REST/CLI, and AWS + Google Cloud from day one. Apache-2.0, 샌드박스 4종 지원, AWS·구글 클라우드 첫날부터.

maybe we should put some AgentEnv environments through EnvCheck!

Open-sourcing the environment layer matters as much as the models, since most agent evals still break on flaky, hand-rolled sandboxes rather than on the agent itself.