AI Engineering Trends: Agent Evaluation, Registry Governance, and MCP Discovery

Today's AI trends

Framework-neutral agent evaluation is becoming a concrete product boundary. AWS’s August 26 update positions AgentCore Evaluations around compatible OpenTelemetry or OpenInference traces rather than a single agent framework. The practical implication is to treat instrumentation as an architectural dependency when choosing or migrating agent runtimes. Source

Agent capability catalogs are moving toward platform governance. AWS made Agent Registry generally available on August 31 as a private catalog for agents, tools, skills, MCP servers, and custom resources, with search, approval workflows, access control, lifecycle tracking, and CloudTrail auditing. For platform teams, reusable agent capabilities now need ownership and approval paths much like internal APIs. Source

MCP is leaning into stateless, HTTP-native tool infrastructure. The July 28 specification removed the protocol initialization handshake and session header, added server/discover, and describes ordinary HTTP-based horizontal scaling; its roadmap highlights progressive discovery as tool catalogs grow. Design tool services for bounded discovery and explicit long-running work rather than assuming a small, permanently loaded tool set. Specification Roadmap