Amazon Web Services launched the Amazon Bedrock AgentCore harness as a generally available service that aims to let teams stand up production-grade agents with two API calls.
The company said the harness bundles existing AgentCore primitives into a managed runtime that handles isolation, persistent memory, a tools gateway, identity vaulting and observability.
CreateHarness defines an agent and InvokeHarness runs it, and new GA behaviors include automatic memory provisioning, a toggle to load an AWS-curated skills catalog without network fetch, and real-time streaming of invocation steps to Amazon CloudWatch.
The harness also lets a session switch models mid-conversation so a team can plan with one model, write code with another and summarize with a third while context persists on AWS.
AWS reported agent tasks on AgentCore grew 15 times in six months and named Nasdaq, Visa and Experian among customers scaling agents in production.
The package increases portability at the model layer but creates operational lock-in because memory, gateway, skills and identity remain AWS services.
AWS offers an export path via a single CLI command that emits Strands-based code, with a Claude Agent SDK export listed as coming soon.
There is no charge for the harness itself, but runtime bills by the second, the gateway by invocation, memory by event and retrieval, and observability at standard CloudWatch rates.
Enterprises should weigh ownership of agent logic versus AWS-managed primitives and model cost predictability under production load.
Microsoft expects its hosted agents with sandboxed sessions and filesystem access to reach general availability by early July.
