LedgerLab LedgerLab
LedgerLab LedgerLab is a memory-first Jupyter workspace environment for training long-horizon business agents with OpenEnv.
Core trick
Agent works in a live Jupyter kernel with reference spreadsheets, and reward is tied to submitted notebook deliverables.
Description
LedgerLab
LedgerLab is a memory-first Jupyter workspace environment for training long-horizon business agents with OpenEnv.
This project targets the OpenEnv hackathon theme of long-horizon instruction following and business workflows. The environment forces an agent to work through realistic spreadsheet and document tasks by inspecting reference files, creating notebooks, running iterative analysis, producing deliverables, and finally submitting for reward.
README excerpt
# LedgerLab LedgerLab is an OpenEnv environment for training notebook-driven business agents on long-horizon professional tasks. It models the real workflows of a financial analyst: reading spreadsheets, building notebooks, iterating on analysis, and producing deliverables — not just answering questions. ## What LedgerLab Models Most agent benchmarks treat the world as a question-answering surface. LedgerLab treats it as a workspace. The agent operates in a partially observable environment with reference files, a live Jupyter kernel, persistent memory, and a reward function tied to concrete deliverables. This design is grounded in how professional knowledge work actually happens: - The analyst does not know in advance which files contain the right data - Intermediate outputs change what steps come next - Mistakes must be detected and corrected within the same session - The final deliverable must be consistent with everything computed along the way ### The Agent's World The agent starts each episode with a workspace containing business documents — spreadsheets, ledgers, inventory files, lease records, location data. It must build an accurate internal model of that workspace through tool use, then produce a verified output. Key world-modeling demands: - **Partial observability**: the agent must discover relevant data by inspecting files, not by being told what they contain - **State accumulation**: information gathered in early steps must be carried forward correctly th