Skill bundles

Run broad capability evals before and after RL training to catch degradation

Detect when RL training narrows capability (great on trained tasks, worse on everything else)

Quantify how much RL training on coding transfers to (say) data analysis or writing

Build evals that test whether RL training on task A improved performance on related task B

Classify agent skills by risk level (read-only vs. write vs. financial vs. external-facing) and apply appropriate controls

Log every decision an RL agent makes in production with sufficient context for post-hoc review

Build eval-gated deployment pipelines where RL-trained models must pass benchmarks before production

Monitor and prevent agents from leaking sensitive data through tool calls

Static and dynamic analysis of agent skill code for security vulnerabilities

Test whether agents behave differently when they believe they're being evaluated vs. not

Verify that the policy optimizes for the intended objective, not a proxy

Restrict agent actions to prevent irreversible or harmful operations