Today's AI trends
Agent evaluations move from answers to end states. NVIDIA’s September 21 guidance recommends executing tool calls in an environment, tracking state across steps, and separately scoring process quality and final task completion. The practical implication is to gate agent releases on verified postconditions and retain traces for diagnosis, rather than relying on a polished final response. NVIDIA Technical Blog
Code-running agents are being treated as constrained workloads. An AWS case study says Benchling combines AgentCore Code Interpreter in VPC mode with Route 53 Resolver DNS Firewall and VPC endpoint policies to reduce data-exfiltration paths; it reports more than 600 code-execution sessions per day and more than 250 distinct tenants per week after deployment. The practical implication is that process isolation alone is not enough: explicitly design network egress, DNS, endpoint access, credentials, and tenant boundaries. AWS Machine Learning Blog