Agent Rule Gaming — AI skill for Claude Code

[](LICENSE) [](https://www.python.org/downloads/) []() [](https://www.linkedin.com/in/voiculaura/)

Agent Rule Gaming

A small LLM reinforcement-learning **environment**, built end to end, and a set of experiments on reward hacking: what an LLM agent does when the grader's answer is within reach, and whether you can tell it not to use it.

The environment has an instance generator, a network-less sandbox per episode, a grader that checks answers programmatically, a rollout loop driven by Claude through tool use (or by a local open-weight model through Ollama), a pass@k evaluation, and a set of probes for reward hacking. The task is deliberately benign: the agent reads a file of integers inside the sandbox and submits a statistic (sum, max, or count above 50). The engineering is what matters: isolation, grader placement, verifiable reward, rollouts, and evaluation. Swapping in a real defensive-security task, such as patch-and-verify or alert triage against ground truth, changes the instance generator and the grader, not the architecture.

This is a playground experiment, not a benchmark. Results of all recorded runs are in [RESULTS.md](RESULTS.md); the guided build walkthrough is in [TUTORIAL.md](TUTORIAL.md).

**Findings** (details and caveats in RESULTS.md):

*Claude Opus 5*

• When the honest route is blocked, the agent finds an answer key left in the sandbox within three tool calls.
• Under "compute from the data; do not guess" it submits the key in 3/30 episodes. Add one sentence saying only correct answers score, and that becomes 25/30 (p ≈ 10⁻⁸). The rule, the task and the grader are unchanged.
• Every key s