Wood Mackenzie Builds Shared Agentic Platform APEX on Amazon Bedrock AgentCore

The problem: 88% of AI prototypes never ship

Wood Mackenzie, the energy and natural resources research firm, says 88% of its internal AI proofs-of-concept never reach wide-scale deployment. That figure lines up with broader industry data the company cites: enterprise AI experimentation is near-universal by early 2026, but only about a quarter of organizations have scaled agents into production in even one business function.

According to Wood Mackenzie, the gap is architectural, not a model quality problem. Forrester research cited in the company's account attributes most agent failures to ambiguity, miscoordination, and unpredictable system behavior — not conventional software bugs. Teams also report they cannot reliably predict when a non-deterministic agent will fail, and that they lack a fast way to shut down a misbehaving agent once deployed.

The fix: one shared runtime instead of three

Before building a shared platform, three separate Wood Mackenzie applications — Woody, Lens AI, and the ST Trading App — were each independently standing up their own agent infrastructure: separate runtimes, separate identity systems, separate observability, and separate hardcoded models. The company built APEX (Agentic Platform for Energy eXperience) on Amazon Bedrock AgentCore to consolidate this into a single shared backend.

AgentCore, which reached general availability in October 2025, is AWS's managed platform for running agents built with any framework — including LangGraph, LangChain, LlamaIndex, CrewAI, Google ADK, Strands Agents, and OpenAI's Agents SDK — and any model, whether or not it runs on Amazon Bedrock. It supports both the Model Context Protocol (MCP) and Agent-to-Agent (A2A) protocol.

Wood Mackenzie evaluated AgentCore against LangChain, LangSmith, CrewAI, n8n, and direct frontier-model APIs on hosting, cost model, model agnosticism, scalability, and governance before selecting it. Five capabilities decided the choice, according to the company:

• Managed infrastructure — AWS handles scaling, patching, and availability with no cluster to operate; at GA, AgentCore added support for Amazon VPC, AWS PrivateLink, CloudFormation, and resource tagging.
• Model agnosticism — Claude, GPT-4.1, Amazon Nova, Mistral, and Llama are all reachable through one platform, allowing teams to swap providers without rewriting business logic.
• Automatic scaling — AgentCore Runtime scales from zero to thousands of concurrent agent invocations with full session isolation and execution windows up to 8 hours.
• Native guardrails and identity — content filtering, PII detection, and policy enforcement are managed rather than reimplemented per team; AgentCore Policy converts natural-language rules into Cedar, AWS's open source policy language, and intercepts tool calls in real time via the Gateway. AgentCore Identity adds IAM integration, VPC isolation, and encryption.
• AWS Enterprise Support — 24/7 support plus a technical account manager relationship the company says lets it influence the AgentCore roadmap.

AgentCore bills on consumption, with no upfront commitment or minimum fees. Runtime charges are based on active CPU and memory consumption per second, and CPU does not accrue during I/O wait — a detail Wood Mackenzie flags as significant because agentic workloads reportedly spend 30–70% of their time idle, waiting on model responses, tool calls, or database queries.

Architecture

APEX is structured in layers. End users interact with three applications — Woody, Lens AI, and the ST Trading App — through an APEX frontend framework that lets product teams add agentic features to a UI without rebuilding the backend connection each time. The APEX backend runs on AgentCore, alongside a Wood Mackenzie infrastructure-as-code framework using AWS CDK and GitHub. A companion control plane, APEX Studio, gives teams a single place to operate the shared platform, including identity-aware entitlements that let the same infrastructure serve internal users (Woody) and external customers (Lens) under different access rules.

What this means

This is not a new model release — AgentCore is AWS infrastructure, not a trained model, and Wood Mackenzie's APEX is an internal engineering platform built on top of it. The news value is in the pattern: a large enterprise consolidating three independently-built agent stacks into one shared runtime specifically to solve observability, identity, and governance problems that surveys say block roughly three-quarters of enterprise agent projects from reaching production. The 88% internal failure rate figure is a company claim, not independently verified, but it's consistent with the broader industry numbers cited. For enterprises evaluating build-vs-buy on agent infrastructure, the operative detail is AgentCore's per-second, I/O-wait-excluded billing model and its explicit model-agnosticism — both aimed squarely at the two biggest complaints about self-hosted agent stacks: idle compute cost and vendor lock-in.