2026-10-21
Introducing AgentEnv — and how REA captures feed its environments
On 2026-10-05, Scale Labs open-sourced AgentEnv, an open-source framework for building reinforcement-learning (RL) environments. Scale frames it as the stack every RL environment they build runs on. This page is a community teaching rewrite for REA operators: what AgentEnv is, how REA session captures show up as Hugging Face replay datasets, and what rea.run does (and does not) host. It is not a verbatim reprint of the Scale Labs post.
Attribution: AgentEnv and the HF collection are published by Scale Labs / earakely-scale (Edgar Arakelyan and collaborators). rea.run is an independent community mirror catalog — not affiliated with Scale.
1. What AgentEnv is
AgentEnv is Scale Labs’ open-source RL environment framework: a way to define simulated slices of the real world (tools, data, rubrics/verifiers) so agents can attempt tasks and learn from trajectories. Scale’s public framing positions AgentEnv as the shared substrate under their own RL environments, now available for others to build with.
Two design pillars show up repeatedly in the Scale write-up:
| Pillar | Idea (paraphrased) | Operator takeaway |
|---|---|---|
| Building blocks | An Environment is composed of versioned pieces: the Environment itself (Slack, Gmail, a help desk, Blender, …), an Artifact of data that populates it, and Rules (clock, triggers, roles) that govern behavior | Build the world once; swap data/rules for different tasks |
| Interoperability | Environments do not assume a particular agent or sandbox. Discovery uses an Environment Card (open AgentEnv Environment Protocol, often at /.well-known/agent-env.json); tools can be reached over MCP, REST, or a generated CLI | Pair any Environment Card with any Agent Card (A2A-style) without custom glue |
Tasks are modeled as a DAG of steps (setup world, load data, bring in an agent, grade with a rubric or strict check). AgentEnv ships 48 built-in step types covering common RL task construction; custom steps can be plugins. Sandboxes are pluggable — Scale names Docker, Modal, and E2B among day-one options (plus registering your own).
That is the product surface. The rest of this page answers the REA-specific question: *where do recorded browser sessions fit?*
2. Three layers: REA capture vs AgentEnv replay vs official runtime
Operators often collapse three different jobs into one word — “environment.” Keep them separate:
| Layer | Who | What it does | What rea.run hosts |
|---|---|---|---|
| REA capture | You + rea-agents | Record an authorized browser session / HAR (or related capture path) against a target you own or have written permission for | Install + FAQ docs only — not capture blobs |
| AgentEnv replay / train | Scale AgentEnv + HF publishers (earakely-scale) | Turn captures into replayable MCP tools and RL-style tasks for training and eval | Nothing — runtime stays off this site |
| Official runtime / sandboxes | Scale AgentEnv cloud, Docker/Modal/E2B, HF Space demos | Actually run the environment, grade trajectories, host sandboxes | Nothing — we do not ship parquet, HAR, or an RL runner |
rea.run is the catalog layer only: thick HTML index + detail pages that deep-link to Hugging Face. If a page implies we host the RL runtime, treat that as a bug.
3. The HF collection: ~50 datasets + 1 Space
Scale Labs engineer earakely-scale published the collection REA Environments on AgentEnv: about 50 dataset environments plus a start-here Space (earakely-scale/rea-environments). Catalog notes on this site describe REA-built replays from recorded sessions of self-hosted apps and public APIs, exposed as MCP tools on AgentEnv.
| Resource | URL | On-site mirror |
|---|---|---|
| HF collection (~50 datasets + Space) | earakely-scale/rea-environments-on-agentenv | /environments/ |
| Start-here Space gallery | earakely-scale/rea-environments | Linked from each catalog row |
| Example dataset id style | earakely-scale/GiteaEnv-REA (names stay English) | /environments/GiteaEnv-REA/ |
Browse categories, tool/task counts, and baseline notes on /environments/. Open the HF dataset or Space when you need bytes or a live gallery — Pages never mirrors those files.
Representative environments (on-site details)
Deep-link a few teaching mirrors — each page cites HF note fields and points back to the dataset. Names stay English:
| Environment | What it rehearses | On-site detail |
|---|---|---|
| GiteaEnv-REA | Self-hosted Git, issues, code review | /environments/GiteaEnv-REA/ |
| GrafanaEnv-REA | Observability dashboards, alerts, datasources | /environments/GrafanaEnv-REA/ |
| SentryEnv-REA | Error tracking and issue triage | /environments/SentryEnv-REA/ |
| NocoDBEnv-REA | No-code database (spreadsheet-style) | /environments/NocoDBEnv-REA/ |
| AirbyteEnv-REA | ELT connectors and sync jobs | /environments/AirbyteEnv-REA/ |
| BookStackEnv-REA | Team wiki (shelves, books, pages) | /environments/BookStackEnv-REA/ |
| MattermostEnv-REA | Team chat channels and posts | /environments/MattermostEnv-REA/ |
| KeycloakEnv-REA | Identity / SSO administration | /environments/KeycloakEnv-REA/ |
Full catalog: /environments/.
4. Path checklist
Use this sequence when teaching or filing a ticket:
- Authorize the target — own software, OSS under a compatible license, written permission, or an approved lab. See /compliance/ and /faq/legal-boundary/.
- Capture with REA — install and doctor-green: /docs/install/. Record only what you are allowed to hold.
- Browse the Environments catalog — /environments/ for the community mirror of HF notes (category, tools/tasks, baselines).
- Open on Hugging Face — collection or per-dataset page for the actual artifacts and Space gallery.
- Read Scale’s original — product depth and framework details live on the Scale Labs post (Source below). This page does not replace it.
5. Compliance one-liner
Community mirror / authorized capture only. rea.run lists what earakely-scale published; it does not grant rights to capture third-party apps, redistribute proprietary HARs, or run AgentEnv for you. Capture and replay stay on targets you are allowed to analyze.
FAQ (short)
| Question | Answer |
|---|---|
| Does rea.run run AgentEnv? | No — catalog + teaching pages only. /faq/agentenv-hosting/ |
| Where are the ~50 datasets? | Hugging Face collection; mirrored as HTML on /environments/ |
| Keep English dataset names? | Yes — e.g. GiteaEnv-REA. Chinese UI uses 「环境」 for the site section |
| Primary CTA for “Introducing AgentEnv”? | This on-site post (and the ZH twin). Scale URL is Source only |
Related
- /environments/ — REA → AgentEnv mirror catalog
- /faq/agentenv-rea/ — short FAQ
- /faq/agentenv-hosting/ — we do not host the RL runtime
- /docs/install/ — install REA
- Chinese edition: /zh/blog/agentenv-rea/
Source
Scale Labs original (2026-10-05): Introducing AgentEnv: An Open-Source Framework for Building RL Environments — authors Edgar Arakelyan, Tejas Polakam, Yzabelle Go. Reference points rewritten above (building blocks Environment/Artifact/Rules; Environment Card + MCP/REST/CLI; Tasks as DAG; 48 built-in step types; pluggable sandboxes Docker/Modal/E2B) were verified against that post. HF collection and Space remain the dataset sources of truth.