AWS Strands Agents Team Releases Strands Harness

The AWS Strands Agents team has released Strands Harness, a fully assembled, general-purpose agent harness aimed at simplifying the development and deployment of AI agents. This open-source framework addresses a common challenge developers face where agent ideas that function well in integrated development environments like Claude Code or Codex struggle when rebuilt with custom loops. Strands Harness is designed to run locally or deploy to various cloud providers and is available for Python and TypeScript under the Apache 2.0 license, requiring only a single line of code to initiate.

According to the Strands Agents team, Strands Harness demonstrates a 28% lower cost compared to other harnesses when running the same Claude or GPT models across six benchmark tests, while achieving near-equal accuracy. The framework is designed for deployability, supporting local execution and offering a bundled skills file that assists coding agents in generating deployment configurations for major cloud platforms including AWS, Google Cloud Platform (GCP), Azure, Cloudflare, and Modal. A harness, in this context, refers to the system surrounding the core AI model, encompassing the execution loop, tool integration, context management, memory functions, and error recovery mechanisms.

Strands Harness builds upon the existing Strands Harness SDK, which previously exposed these building blocks. The new harness packages these components into functional default settings. It is engineered as a general-purpose agent, not exclusively a coding agent. The `create_harness()` function, out of the box, provides an agent capable of operating with current reasoning models from various providers such as Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM. It includes essential tools like shell commands, file operations (read, write, edit), and web browsing capabilities, moving away from bespoke tools for each specific task. To manage large tool outputs efficiently, it offloads them to files and caches frequently reused parts of requests. For persistent interactions, it maintains long-term memory across sessions and can resume conversations using a session ID. Complex or open-ended subtasks are delegated to a built-in helper agent, and multi-step workflows are tracked using a checklist. The framework also supports loading additional Agent Skills as they become available.

The reported 28% cost reduction was determined through distributed benchmarking conducted by the Strands Agents team on Amazon EC2 instances, utilizing Harbor, an evaluation framework developed by the creators of Terminal-Bench. The cost metric is an average dollar amount per task across six benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1. Competitors evaluated in this comparison included Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness. The Strands Harness aims to democratize AI agent development by providing a robust, cost-effective, and user-friendly platform for developers to build and deploy sophisticated AI applications.