DeepGym provides infrastructure for reinforcement learning (RL) training and evaluation specifically designed for coding agents. It addresses challenges faced by code model teams, such as building and maintaining sandbox environments, test harnesses, and evaluation pipelines, by offering a ready-made stack that handles these requirements. The platform aims to reduce the time and effort teams spend on infrastructure, enabling them to focus on training and improving AI models.
0. It offers sandboxed execution of real code in Daytona containers with full OS-level isolation, network restrictions, and resource limits, as well as a local mode for development. DeepGym supports adversarial testing through built-in reward hack detection strategies that identify empty solutions, hardcoded results, and pattern exploits, and it can discover novel attacks using RL-based exploit discovery. The platform enables step-by-step agent interaction for multi-turn agents, supports computer-use and tool-use tasks such as browser interaction, screenshot verification, and file system operations, and provides fine-grained, per-test-case rewards with detailed input summaries and error traces.
DeepGym is delivered as a Python package installable via pip, and offers both a command-line interface and a browser-based debugging UI for interactive testing. It includes a REST API with OpenAPI documentation for running episodes, batch scoring, and full evaluation suites, with API key authentication for production use. The system supports asynchronous and batch operations, enabling scalable parallel training runs and mixed benchmark batches. Users can share environments and evaluation results by pushing them to the HuggingFace Hub, and register DeepGym environments as lm-eval tasks for use with the lm-eval CLI.
The platform integrates with several RL and AI frameworks, including HuggingFace TRL, DAPO, verl, and OpenRLHF, providing thin adapters and drop-in reward functions for seamless use within existing training pipelines. It offers a Gymnasium-compatible API, allowing it to work with any RL framework that supports the Gym standard. DeepGym is positioned as an RL training and evaluation stack tailored for code model development, targeting teams and researchers building AI coding agents.
deepgym sits in PulseGate's AI & ML category. It facilitates reinforcement learning research by providing reliable training environments for coding agents. It is built as an open-source project for AI researchers and developers. deepgym is open source under the MIT license. The product ships for the web and the command line.
Behind deepgym is abhishekgahlot2, and the product first shipped in 2026. PulseGate's similarity index finds few close equivalents — deepgym occupies a relatively distinct niche. Key capabilities include RL environments, verifiable rewards, and coding agent support.
Latest indexed changes and source events
Other apps tracked under the same category.
