Building an evaluation feedback loop so a coding agent can tune game difficulty
A developer runs headless game simulations and wants a coding agent to adjust difficulty parameters from run metrics. The suggested pattern: feed the agent a small JSON summary (seed, params, 3-5 metrics) rather than raw logs, keep a holdout set of seeds the agent never tunes against, and cap iterations — otherwise the agent 'improves' difficulty into unplayability.