Every agent has an orchestration layer which contains the loop that calls the model, decides which tool to invoke, passes results back, manages context windows, and handles failures. Running that loop requires infrastructure underneath it: compute to host the agent, a sandbox to safely execute code, secure connections to tools, persistent storage, and error recovery. This infrastructure forms the agent harness, enabling an agent to actually run.

Until now, building that harness was the first thing every team had to do from scratch. That meant choosing a framework, writing the orchestration code, connecting it to tools and memory, and ensuring authentication, all before the agent could process a single request. It’s necessary work, but it’s not the work that tells you whether your agent is going to be useful. Most teams we’ve worked with spent days on this infrastructure before they could run their first real test.

The new managed agent harness feature in AgentCore replaces all that upfront build with a straightforward configuration. You declare your agent and run it in just three API calls, without writing orchestration code. You define what your agent does: which model it uses, which tools it can call, and what instructions it follows. AgentCore’s harness stitches together compute, tooling, memory, identity, and security to create a running agent that you can test in minutes. Trying a different model or adding a tool is a config change, not a code rewrite. You can test several variations of an agent in minutes by changing the API parameter on the fly.

That speed doesn’t come at the cost of flexibility. The harness in AgentCore is powered by Strands Agents, the open source framework from AWS. When you need custom orchestration logic, specialized routing, or multi-agent coordination, you switch from config to code-defined harness, with the same platform, same microVM isolation, same deployment pipeline. AgentCore persists session state to a durable filesystem, so agents can suspend mid-task and resume exactly where they left off. This makes human-in-the-loop patterns practical without custom plumbing, and without redesigning the agent later when those needs arise. You can get started in minutes then add more capabilities and control when your needs evolve, without any rearchitecture.