This is my personal repo for RL environment engineering. I’m iterating on an agent that interacts with a custom RL-style environment and uses Anthropic tools to adapt behavior over time.

Project goals

• The initial policy/agent should intentionally fail about 60% of the time (i.e., succeed roughly 40%) to ensure the environment is challenging and informative.
• Over repeated interactions, the agent should learn and demonstrate improvement, achieving the targeted behavior with number of repetitions.

Prerequisites

• • Git (latest)
• Python 3.11+ (see .python-version for the exact version)
• uv for Python project management (pip install uv or see https://docs.astral.sh/uv/)
• An Anthropic API key

Setup (Windows PowerShell)

1. Clone and enter the repo

git clone <repo-url>

2. Install dependencies with uv

uv sync

3. Set your Anthropic API key in the current shell

$env:ANTHROPIC_API_KEY="your_api_key_here"

4. Run the agent

uv run main.py

Execution Modes

The agent can run tests concurrently or sequentially. Change the concurrent flag at the bottom of main.py:

asyncio.run(main(concurrent=True))
asyncio.run(main(concurrent=False))

When running concurrently, results print as they complete (not in run order) for faster overall execution.

Benchmark reference

This project references and draws inspiration from large-scale software engineering benchmarks such as SWE-Bench Pro. For details about the dataset, task structure, and evaluation setup, see the dataset page on Hugging Face: ScaleAI/SWE-bench_Pro.