Amazon Bedrock AgentCore Consulting

What AgentCore actually gives you

Every team that ships an agent ends up building the same six things: somewhere isolated to run it, a governed way to reach tools, memory that survives the session, an identity that travels with the request, traces you can read when it misbehaves, and a way to tell whether it is any good.

You can build all six. Teams do. The problem is that you then own them — forever, alongside the product you actually meant to build.

Amazon Bedrock AgentCore provides them as managed services: Runtime, Gateway, Memory, Identity, Observability and Evaluations, plus a Managed Harness and optional Payments. It is bring-your-own-framework, so Strands, LangGraph, or three sequential prompts with an explicit control loop all sit on top of it unchanged.

The lifecycle change worth knowing about

On June 30, 2026 AWS renamed the original Bedrock Agents — the November 2023 service — to Bedrock Agents Classic. It entered maintenance for new customers after July 30, 2026.

Existing Classic deployments continue to operate, so this is not an emergency. But it does mean two things. Net-new agent builds should start on AgentCore. And teams currently running Classic agents should have a deliberate plan rather than discovering the situation during an incident.

Where the boundary lives

The most common production failure in agent systems is not a bad model output. It is an agent with more write access than anyone intended, discovered after it used it.

Gateway is the answer to that. Every tool an agent can reach passes through it, with customer-configurable rate limits and Cedar-based fine-grained access control on writes — extended to Memory namespaces on August 28, 2026. The practical consequence: mutating an order, a price, or a customer record becomes a permission decision evaluated outside the model. A model can be argued out of a system prompt. It cannot be argued out of an authorization policy.

Identity propagation is the other half. When the acting user identity travels with the request, instead of every tool call arriving as one over-privileged service role, the audit trail becomes something a compliance reviewer can actually use.

Evals are not optional

An agent without an eval suite is a system whose quality you are asserting rather than measuring.

We build a golden dataset per agent during development and set a pass bar it must clear before deployment. AgentCore Evaluations runs it on every deploy. If it does not clear, it does not ship — the sprint calendar does not get a vote. This is the single practice that most reliably separates agents that are still running in ninety days from agents that were quietly switched off.

Modelling the bill honestly

AgentCore bills across several components rather than one meter, and first estimates are usually wrong in the same direction — too low, because the model inference line dominates and tool-call volume per conversation is underestimated.

We publish a free Amazon Bedrock AgentCore pricing calculator covering the components. It is the same one we use on the first call, and we would rather you ran it before talking to us.

What we do

We design the AgentCore stack against your actual workflow, deploy it as infrastructure-as-code in your accounts, wire the Gateway boundary and Cedar policies, build the eval harness and the golden dataset, instrument cost ceilings and CloudWatch alarms, and hand over the runbooks.

Your engineers are in the pull requests throughout. There is no proprietary orchestration layer to be locked into — the whole point of AgentCore is that the substrate is AWS-managed and the agent logic is yours.

For the commerce-specific application of all of this, see eCommerce AI Agents on AWS. For the broader Bedrock platform — Knowledge Bases, model selection, Guardrails — see Amazon Bedrock Consulting.