Wood Mackenzie has built a common platform for deploying AI agents so product teams do not have to recreate authentication, memory, monitoring and safety controls for every application. Called APEX, the system supports both employee tools and customer-facing products on Amazon Bedrock AgentCore.
Before APEX, three applications were developing separate agent stacks. That would have duplicated infrastructure and made it harder to share tools, evaluations or conversation state. The new design puts runtime services, identity-aware access and observability in one managed layer while leaving each team responsible for its business logic.
The company says teams can use frameworks including LangGraph, LangChain, LlamaIndex, CrewAI, Google ADK, Strands and OpenAI’s Agents SDK. Models from several providers can also be swapped without rewriting the surrounding application. Support for Model Context Protocol and Agent-to-Agent connections gives the platform standard interfaces for tools and cooperating agents.
This is a customer case study from AWS, not an independent comparison with self-hosted alternatives. Centralization also creates a shared dependency whose governance and reliability need careful management. The architectural lesson is still useful: enterprise agent projects often fail on session isolation, permissions and evaluation rather than model access, so those controls benefit from being treated as platform capabilities instead of repeated application code.