Operate AI agents like mission critical infrastructure.

Continuing from Part 1, this post explains how identity, delegation, capability registries, and human-approval workflows govern MCP capabilities. It shows why GUI users, applications, and AI agents need different discovery, risk, and approval treatment—while remaining consistent through one enterprise control plane.

This opening article in a three-part series explains why MCP security must move from protecting API endpoints to governing capabilities. It covers caller-aware identity and authorization for humans, applications, and AI agents, and why the same action can require different trust thresholds depending on who is calling.

As AI agents become the primary operators of software, this post argues that MCP can become the universal interface. GUIs shift toward supervision and approval, while agents consume capabilities rather than screens—making MCP a stronger front door than REST for discovery, execution, and governance.

As MCP adoption grows, this post explains why traditional API gateways are not enough for production-grade agent systems. It outlines how MCP-aware gateways enable tool-intent routing, semantic caching, persistent session handling, tool-level authorization, and actionable observability that better match stateful, multi-agent MCP workflows.

As MCP ecosystems shift toward dynamic, runtime-discovered agent architectures, this post explains why CMID and DCR should move out of individual agents and into the MCP Gateway. Centralizing identity handling reduces operational overhead, prevents identity sprawl, enforces policy consistently, and creates a more stable trust boundary across rapidly changing multi-agent systems.

In this final post in the mini-series on Agent evals, the focus shifts to qualitative evaluation using rubric-based metrics across RAG agents, general chatbots, and specialized agents. It covers how LLM-as-a-judge scoring can be aligned with thresholds and practical planning considerations such as retrieval quality, reranking, and metadata labeling for robust real-world evaluation.

Building on agent architecture foundations, this post dives deeper into testing and evaluation strategies that differ across agent types. We explore LLM structured output testing for specific use-cases like SQL query generation, traditional structured testing for planner-executor models, and generative outcome assessment using LLM-as-a-judge with qualitative rubrics. Learn how to customize evaluation approaches and select the right metrics for real-world agent deployments.

Enterprise business logic has traditionally been deterministic and easily testable, but AI agents introduce new challenges in evaluating reliability and safety. This post explores different architectures for AI integration—from hybrid models with deterministic cores to planner-executor approaches—and the evaluation strategies needed for each to help enterprises confidently migrate to AI-assisted workflows in production.