A shared agentic platform for Wood Mackenzie, on Amazon Bedrock AgentCore

Creating a functional agentic prototype can be completed in a single afternoon; however, scaling it for production presents significant challenges. When an agent must serve multiple users, additional engineering requirements emerge, making it essential to determine if the agent performs correctly under real-world conditions. Concurrency, session isolation, identity, persistent storage, scaling, and safety measures constitute layers that most teams must construct independently, despite the benefit of a standardized platform for agents.

As a consequence, a considerable gap exists between experimentation and production. According to industry surveys conducted up to early 2026, enterprise AI experimentation remains widespread, yet only approximately a quarter of organizations have successfully deployed agents in production across any function. Within Wood Mackenzie, research indicates that 88 percent of AI proof-of-concepts fail to achieve widespread deployment.

The primary reasons for this issue are architectural in nature, rather than associated with the models themselves. Forrester attributes agent failures predominantly to ambiguity, miscoordination, and unpredictable system behavior, rather than typical software bugs. The most frequently mentioned obstacle is evaluation and observability — teams are unable to reliably determine in advance if a non-deterministic agent will produce incorrect results, and standard regression tests are insufficient to detect these issues.

Governance and identity are closely related concerns, with a substantial number of executives reporting that they are unable to promptly halt a malfunctioning agent. At the core of this problem lies duplicated infrastructure: each team independently implements authentication, safety mechanisms, memory management, and tracing, and must manually hardcode a specific model, making it necessary to rewrite code when switching providers.

This scenario exemplifies the inefficiencies associated with building a shared platform. The solution lies in a unified agentic platform that abstracts away the common infrastructure, enabling teams to focus on business logic unique to their products. The APEX platform, developed by Wood Mackenzie, addresses this need. Built upon Amazon Bedrock AgentCore, APEX offers a shared runtime that handles orchestration, safety, observability, identity, and connectivity, thereby eliminating the necessity for teams to recreate these foundational elements.

AgentCore is a platform engineered for large-scale agentic deployments, providing a model-agnostic environment that can accommodate any framework or model. It addresses key requirements for a production-ready platform, including hosting models, cost management, flexibility regarding the choice of model, scalability, governance, and enterprise-level support.

In evaluating AgentCore against these requirements, it was determined that it offers a managed platform rather than a library that would need to be maintained internally. The platform is compatible with a variety of open-source frameworks, such as LangChain, Strands Agents, LangGraph, LlamaIndex, CrewAI, Google ADK, and OpenAI Agents SDK, as well as any model, regardless of whether it runs on Amazon Bedrock.

It supports both the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) protocol. This compatibility allows for standardization of the platform layer while preserving the flexibility to select individual frameworks and models for each team. The five capabilities that led to the adoption of AgentCore include managed infrastructure, true model agnosticism, automatic scaling, native guardrails, and enterprise-level support.

Managed infrastructure eliminates the need for managing clusters, as AWS handles scaling, patching, and availability, allowing the platform team to concentrate on agent capabilities instead of backend operations. AgentCore was made generally available (GA) in October 2025 and includes support for Amazon Virtual Private Cloud (VPC), AWS PrivateLink, AWS CloudFormation, and resource tagging from its inception.

It supports a wide array of models, including Claude, GPT-4.1, Amazon Nova, Mistral, and Llama, enabling teams to plan with one model and execute with another, facilitating price-performance testing by swapping providers without losing conversation context or altering business logic. Automatic scaling is facilitated by the AgentCore runtime, which can scale from zero to thousands of concurrent agent invocations without requiring any capacity planning, while maintaining complete session isolation and execution windows of up to 8 hours.

Native guardrails and identity management are provided as managed capabilities, automating content filtering, personally identifiable information (PII) detection, and policy enforcement through integration with the Gateway. This approach simplifies the process of building and managing secure, compliant agents. The platform's architecture is designed to serve both internal users, such as Woody, and external consumers, like Lens, by providing identity-aware entitlements and ensuring seamless integration with real-world applications.

As multi-agent systems become the standard, the APEX platform is positioned to evolve to meet these growing demands.