Measuring Biological Capabilities and Risks of AI Agents: Generating and Interpreting Evidence from Agentic Evaluations