Back to Project
F

FDE-Gym

Project Oct 16, 2026

A benchmark and practice platform for agents delivering working systems under ambiguous customer requirements.

Visit primary link 3 sources attached
FDE-Gym authors' promotional workbench image showing a simulated campus booking task, agent panel and illustrative score and business value.
Author-published practice-platform illustration. The score of 99 and $10,000 value shown are simulated promotional example fields, not benchmark aggregate results or real customer revenue. Resized variants preserve the full image. Source: DeepLethe / FDE-Gym authors
agent-ecosystemai-agentsverification

What happened lately

FDE-Gym timeline

FDE-Gym evaluates end-to-end customer delivery in simulated environments with stakeholders, modifiable systems and hidden constraints. Its October 15, 2026 paper describes 95 tasks derived from 20 scenarios across 18 sectors.

The repository provides an Apache-2.0 application and Python harness. Three complete scenarios are publicly available under separate CC-BY-NC-4.0 dataset terms; the original tasks for all 20 scenarios require access approval, and 75 variants are held out for leaderboard scoring. A local Docker setup requires Python 3.9 or later and an OpenRouter key for simulated stakeholders.