Amazon Bedrock AgentCore harness reaches general availability

AWS announced general availability of the Amazon Bedrock AgentCore harness, a managed runtime that compresses the path from idea to production-grade agent into two API calls: CreateHarness to define an agent and InvokeHarness to run it. Each agent runs in an isolated environment with filesystem and shell access and persists memory across sessions — a deliberate move to make stateful, long-running agents a default capability rather than a bespoke engineering project.

The filesystem-and-shell sandbox plus cross-session memory targets the most common pain points teams hit building agents in-house: environment isolation for safety, and persistence so an agent's work today informs its work tomorrow. It dovetails with the broader AgentCore expansion announced at Summit NY — web search, managed knowledge bases, and policy guardrails — positioning AgentCore as the connective tissue between models on Bedrock and real enterprise workflows.

Competitively, this is AWS's most direct answer to Microsoft's Copilot Cowork (now GA worldwide with multi-model support) and to the agent-harness pattern LlamaIndex's Jerry Liu and others have championed. The 'two API calls' pitch echoes the developer-experience simplicity that made Lambda and Bedrock land — abstract away the plumbing, charge for the runtime.

The caveat is cost and lock-in: isolated environments with persistent memory and shell access are powerful but consume compute, feeding the same 'tokenmaxxing' bill anxiety driving Microsoft to explore cheaper models like DeepSeek V4. Enterprises will weigh AgentCore's convenience against multi-cloud portability. Watch whether developers adopt the harness primitive directly or keep building on open frameworks for vendor neutrality.