Building Agents on AWS Just Got a Lot Easier

Building an agent on AWS today means picking a framework, writing the orchestration loop, packaging it as a Docker image or code bundle, and wiring up your own state, memory, and tracing. AgentCore (since October 2025) already solved a lot of that: managed runtime, memory, gateway, identity, observability. But one thing stayed on you: writing the agent code yourself.

The harness removes that last piece. It sits one layer above Runtime. No loop to write, no container to push. You describe the agent via the model, system prompt and tools and AWS runs the orchestration on top! Under the hood it's powered by Strands Agents, but you don't need to know that to use it.

What you get out of the box

A per-session microVM with its own filesystem and shell. Multi-provider models: any Bedrock model, OpenAI, or Google Gemini, with provider switching mid-session. Server-side memory keyed by runtimeSessionId. Built-in tools (browser, code interpreter, MCP, Gateway). And a response stream that is structured observability! You see why the agent called a tool, the params it sent, and what came back, no tracing SDK bolted on.

The 20-line agent

Once the harness is configured in the console, invoking it is tiny. You call InvokeHarnessCommand with the harness ARN, a runtimeSessionId, and the current user message, then read text chunks off the stream. No Lambda packaging, no Docker image, no DynamoDB schema, no conversation history sent by hand. The harness threads it server-side. The session ID is the only state you keep.

The full post walks through configuring a harness in the console, the complete SDK example, and the pricing table line by line.