A shared agentic platform for Wood Mackenzie, on Amazon Bedrock AgentCore
Amazon Web Services Shridhar Navanageri
Wood Mackenzie built APEX on Amazon Bedrock AgentCore so its teams can ship agents without rebuilding the same plumbing. The point is simple: one shared stack for identity, guardrails, scaling, and observability.
Based on reporting by Amazon Web Services, Shridhar Navanageri — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Building an agent prototype is the easy part. The painful part starts when more than one team wants to use it, because then you need identity, session isolation, scaling, memory, guardrails, tracing, and a way to tell when the thing is quietly going wrong in production.
Wood Mackenzie says 88 percent of its AI proofs of concept never make it to widescale deployment. That lines up with the broader problem the company describes: evaluation and observability are still weak, and standard regression tests don’t catch the weird failures that come with non-deterministic agents. So instead of letting each product team rebuild its own stack, Wood Mackenzie created APEX, a shared agentic platform built on Amazon Bedrock AgentCore.
The pitch for AgentCore is not subtle. It is managed, model agnostic, and works with open source frameworks like Strands Agents, LangGraph, LangChain, LlamaIndex, CrewAI, Google ADK, and the OpenAI Agents SDK. It also supports MCP and A2A, and at GA in October 2025 it added VPC, PrivateLink, CloudFormation, and resource tagging. Wood Mackenzie says that combination let it standardize the platform layer without forcing every team onto the same framework or model.
That matters because before APEX, three applications were heading toward three separate agent stacks: Woody, Lens AI, and the ST Trading App. Left to themselves, they would have each stood up their own runtime, observability, identity, and model wiring. APEX instead puts the work into one shared backend with AgentCore Runtime, Identity, Gateway, Memory, Observability, plus an orchestrator, a vector database, Amazon Bedrock model access, and Amazon Bedrock Guardrails.
The interesting part is how much of the platform is about control, not flash. Identity and entitlements are checked throughout the request path, not once at the edge. The Woodmac Agent Registry is there to let one team reuse another team’s agents and tools without copying code. AgentCore Observability pushes telemetry into CloudWatch, and AgentCore Evaluations scores things like helpfulness, tool selection, and accuracy against real traffic. That is the unglamorous work that decides whether agents stay in demo land or become software people actually trust.
My take — AI-written commentary, not fact-checked reporting
This is the right move, and it’s basically the only sane one once agents leave the lab. Everyone loves the demo; nobody loves rebuilding auth, tracing, memory, and policy for the third time because a team wanted a different wrapper around the same mess. The industry keeps calling this “agentic,” but most of the hard part is just platform engineering with a fancier name.
Read more about this at: Amazon Web Services