Amazon Bedrock AgentCore

Amazon Bedrock AgentCore is AWS's attempt to answer a question a lot of engineering teams have been stuck on for the past couple of years: once you've prototyped an agent, where does it actually live? The pitch is a set of 13 composable services — Runtime, Gateway, Memory, Browser, Code Interpreter, Identity, Observability and Harness among them — that you can adopt wholesale or cherry-pick, with AWS absorbing the infrastructure burden underneath.

What stands out immediately is the refusal to be opinionated about the top of the stack. AgentCore is pitched as framework-agnostic and model-agnostic, with explicit support for CrewAI, LangGraph, LlamaIndex, Google ADK and Strands Agents, and the Harness component reaching beyond Bedrock to OpenAI, Google Gemini and OpenAI-compatible providers. For a first-party AWS service, that is a notably un-defensive posture, and it means the platform is plausible even for teams whose model choices are still in flux.

This is squarely a code-first product for platform and backend engineers, not a low-code canvas for business users. The natural audience is an organisation that already runs on AWS, has agent prototypes working locally, and now needs session isolation, OAuth-based tool access, long-running execution and audit trails before anything touches a customer. Teams looking for a visual builder or a batteries-included vertical agent should look elsewhere.

Judged on what AWS and independent coverage describe, the architecture is the strongest part. Turning existing REST APIs and Lambda functions into MCP tools via Gateway is the kind of pragmatic plumbing that saves weeks of bespoke adapter code, and the semantic tool selection to trim context size suggests someone has actually felt the pain of a 60-tool prompt. Likewise, Runtime's per-second billing on real CPU and memory consumption, with CPU scaling to zero during I/O waits and idle memory reclaimed after 120 seconds in v2, is a genuinely well-matched billing model for agents that spend most of their wall-clock time waiting on a model or an API. The eight-hour run ceiling, extendable by configuration, puts it in a different category from request/response function platforms.

The reservations are mostly about maturity and transparency. Independent coverage as of late October 2025 had the platform in preview, and a committed-baseline discount for microVMs is signposted only for October 2026 — both of which point to a product still settling. More awkwardly for a review, there are no published per-unit rates in the material here: "consumption-based with no upfront commitments" is a philosophy, not a quote. With 13 separately metered components, plus ECR or S3 Standard charges for code artifacts and EC2 costs-plus-a-management-fee for Instances, the surface area for surprise line items is large, and the $200 of Free Tier credit for new accounts won't take a serious evaluation very far. Granular metering is a virtue right up until you need to forecast a quarterly bill.

There is also an inherent lock-in question. Modularity means you can take just Memory or just Gateway, which is reassuring, but the deeper you lean on Identity, Observability and Harness together, the more the operational shape of your agent becomes AWS-shaped, even if the model and framework remain portable.

I'd rate this a strong, credible platform with a billing story that is cleverly designed but hard to evaluate. For an AWS-native team, it is the obvious first thing to try; for anyone needing a firm cost model before committing, it warrants a careful, instrumented pilot first. Four stars, with the missing point held back for pricing opacity and preview-era rough edges rather than for any flaw in the design.