HarFeast
HarFeast is a procedurally generated RL environment that simulates real management consulting engagements for a food manufacturing company.
Core trick
Procedurally generated enterprise data worlds expose eight file and compute tools, and each task is graded against deterministic numeric ground truth.
Description
HarFeast is a procedurally generated RL environment that simulates real management consulting engagements for a food manufacturing company. Each "world" is a complete enterprise data ecosystem containing employee surveys (3000+ rows), equipment maintenance records, production data, interview transcripts, scrap rate reports, and industry benchmark documents.
The agent interacts through 8 tools — files.list, files.read, spreadsheet.read_range, data.filter, data.group_by, data.add_columns, data.compute, and submit — that mirror how a real analyst would navigate an enterprise data environment.
Each world contains 14 consulting tasks spanning workforce readiness analysis, equipment failure prediction, training ROI calculation, and operational efficiency assessment. Tasks are scored against multi-criteria rubrics with verifiable ground truth — the agent must produce specific numbers, percentages, and counts that match deterministic computations over the data.
17shasvatj/harfeast_apex_openenv_hackathon★ 01 forksPythonCreated Mar 8, 2026 · last push Sep 9, 2026
README excerpt
# HarFeast **OpenEnv Hackathon 2026 — Mercor APEX-Agents + Scaler AI Labs Enterprise Workflows** A synthetic RL environment where LLM agents learn to do management consulting work: explore enterprise data, use analytical tools, and produce verifiable deliverables. **Live environment:** [openenv-community/harfeast-env](https://huggingface.co/spaces/openenv-community/harfeast-env) --- ## The Problem Mercor's [APEX-Agents benchmark](https://arxiv.org/abs/2601.14242) ([dataset](https://huggingface.co/datasets/mercor/apex-agents)) gave us something very important — proof that AI agents can struggle with performing professional tasks that require multi-step reasoning and tool calling over long horizons. The best models struggle with tasks that a junior analyst handles in an afternoon: navigating fragmented data, chaining analytical steps where each depends on the last, and producing answers with exact numbers that hold up to verification. The problem isn't knowledge — it's workflow. Current RL environments for tool use are too simple: single API calls, obvious databases, one-step answers. Real enterprise work requires sustained reasoning across messy, interconnected data systems. We need environments that teach agents this workflow, and we need to measure whether they actually learn it. HarFeast is that environment. --- ## The Synthetic World At the heart of HarFeast is a **procedural world generator** that creates realistic enterprise data ecosystems. Each world simulat