SpeedRL
APEX Labs trains LLM agents to do real professional work — not pass tests.
Core trick
Files are corrupted with planted errors and an expert gold standard scores each response, while difficulty escalates with agent performance.
Description
APEX Labs trains LLM agents to do real professional work — not pass tests.
We built Crucible: a partially observable RL environment across 22 enterprise scenarios in investment banking, consulting, and law. The environment adapts to the agent — difficulty escalates as performance improves, files are corrupted with planted errors, and every response is judged against an expert gold standard. Agents earn reward for citing evidence, detecting inconsistencies, and outperforming the reference. No shortcuts. Just reasoning.
README excerpt
# APEX Environment **APEX** (Advanced Professional EXperience) is a reinforcement learning environment for training LLM agents on expert-level professional tasks across investment banking, management consulting, and corporate law. Built on the [OpenENV](https://github.com/openenv) framework for the OpenENV Hackathon SF 2026. --- ## What It Is APEX exposes a standard OpenENV interface (`reset()` / `step()`) backed by 22 hand-crafted professional scenarios. Each scenario gives the agent a workspace of real-looking files — financial models, legal agreements, client emails — and a task. The agent produces a professional response. The environment scores it. The environment is not a static benchmark. It adapts to the agent: difficulty escalates as performance improves, scenarios are adversarially corrupted on hard episodes, and every response is scored relative to a gold reference output — not just against a fixed checklist. --- ## Scenarios 22 scenarios across three domains, organised by difficulty tier: | Domain | Easy | Medium | Hard | Total | |---|---|---|---|---| | Investment Banking | 2 | 2 | 4 | 8 | | Management Consulting | 2 | 2 | 3 | 7 | | Corporate Law | 2 | 3 | 2 | 7 | | **Total** | **6** | **9** | **7** | **22** | Each scenario contains: - **`files`** — 3–4 workspace files (financials, memos, emails, policy excerpts) - **`task`** — the specific instruction to the agent - **`rubric`** — 3–7 binary criteria the judge evaluates against - **`gold_output`** — a re