AWS gives enterprise AI agents a shared platform instead of three separate plumbing jobs

AWS has published a detailed account of how Wood Mackenzie built a shared platform for its internal and customer-facing AI agents. The important bit is architectural: one managed foundation handles identity, scaling, memory, observability and guardrails, so each team does not have to rebuild the same machinery before its agent can meet real users.

AWS AI Watch analysis

What happened

Wood Mackenzie’s APEX platform supports three applications, including Woody, Lens AI and the ST Trading App. The AWS account says they share an AgentCore-based backend while retaining freedom to use different frameworks and models. The platform combines AgentCore Runtime, Identity, Gateway, Memory and Observability with Amazon Bedrock models, retrieval, policy controls and an agent registry.

That means an agent can carry a user’s permissions through downstream tool and data calls, keep context across sessions, expose approved tools through a common gateway and send telemetry to CloudWatch. AWS says the runtime can scale from zero to thousands of concurrent invocations, with session isolation and execution windows of up to eight hours. Those are the capabilities AWS describes, not an independent performance test.

Key findings

• Shared foundation
  APEX gives several applications one platform for runtime, identity, memory, tools and monitoring.
• Model flexibility
  Teams can use different frameworks and models without rewriting the business logic around them.
• Continuous oversight
  Observability and evaluations are designed to show when a non-deterministic agent starts behaving badly.
• Permission-aware actions
  Identity and entitlements follow the agent through tool and data calls, rather than stopping at the front door.

Why it matters

The hard part of enterprise agents is increasingly less “can we make a demo?” and more “can we run this for many people without losing track of who may do what?” AWS’s source says Wood Mackenzie found that 88 per cent of its AI proofs of concept never reached widespread deployment, with duplicated infrastructure and weak evaluation among the obstacles.

A shared platform could reduce that repetition. It also creates a new dependency: teams may trade several small stacks for one large control plane, making its pricing, portability and failure modes rather important details rather than footnotes.

Our read

This is a useful production pattern, not proof that AgentCore makes agents reliable by itself. The sensible takeaway is to standardise the boring foundations early, while keeping models, tools and business logic replaceable. Nobody wants to discover their grand AI strategy is three authentication systems in a trench coat.

What to watch

• Whether Wood Mackenzie publishes measured results for reliability, cost and deployment time.
• How easily teams can move agents and data away from the AWS stack.
• Whether shared memory and identity controls behave cleanly across multiple applications.
• How evaluations catch failures that ordinary regression tests miss.

Discussion spark: Should enterprise teams standardise agent infrastructure on one managed platform, or keep separate stacks to preserve portability and limit the blast radius of a failure?

not affiliated with or endorsed by Amazon Web Services (AWS)