AWS Releases 'Strands Harness' Based on Strands Agents SDK with Complete Agent Features Excluding Model Intelligence

AWS has released Strands Harness as an extension of the Strands Agents SDK. This framework is equipped with core features necessary for agent operations, such as web search, command execution, file editing, long-term memory, and helper agent delegation, excluding the model's intelligence.

Performance and Cost Efficiency

AWS claims to have achieved a 26% cost reduction across 7 benchmarks using the same model. In particular, costs were reduced by 77% compared to Claude Code in a specific evaluation, and it recorded high scores on Terminal-Bench 2.1. However, it was also noted that DeepSeek's harness is superior to the Strands harness in terms of token efficiency.

Technical Features and Design Philosophy

Marc Brooker, AWS VP, explained that while the SDK provides building blocks, the harness focuses on solving engineering problems such as context management, tool integration, and system prompt tuning. The main technical challenges and solutions are as follows:

• Context Management: Implemented context compaction solutions to prevent context window saturation from long tasks, such as maintaining relevant history, summarizing past turns, and storing references to large tool results.
• Design Flexibility: After the initial 'tested defaults' design raised concerns about being a black box, it was changed to a pre-configured instance approach in the Strands SDK, exposing all defaults as regular code to allow modification.

Deployment and Ecosystem

Strands Harness supports Amazon Bedrock, Anthropic, OpenAI, Google's latest models, and local Ollama, and is also compatible with Anthropic's Agent Skills format. It can be installed with a single command via PyPI (Python), npm (TypeScript), and CLI, and AWS infrastructure or Amazon Bedrock are not mandatory prerequisites. It can run in various environments, including laptops, CI pipelines, and private servers.

Developers can focus on defining agent goals, connecting systems, and evaluating use cases, and can prototype agents in natural language via the Strands CLI. This product competes with Google's Antigravity agent and Microsoft's Agent Framework Harness, helping developers build high-efficiency agents without becoming harness experts.