The AWS Strands Agents team has released Strands Harness, an open-source framework designed to optimize agent performance. The system provides tools for context management, memory, and recovery, aiming to reduce token costs while maintaining accuracy. It supports local execution and generates deployment configurations for AWS, GCP, Azure, Cloudflare, and Modal.
Benchmarking conducted on Amazon EC2 across six frameworks, including ALFWorld and Terminal-Bench 2.1, shows significant efficiency gains. Compared to Claude Code, Strands Harness demonstrated a 77% reduction in costs while achieving a 7.9-point higher accuracy score. The system also outperformed Oh-my-pi in cost-efficiency while matching its accuracy.
The framework includes built-in defaults for prompt caching and context management, which the team identifies as key drivers for token efficiency. While DeepSeek Harness remains 14% cheaper, it consistently scores lower across all benchmarks. The highest performance recorded was with Claude Opus 5 on Strands Harness, reaching approximately 85% accuracy.
Developers can install the tool via pip or npm and integrate it with local Ollama models. The Strands CLI allows for agent prototyping in plain English, with the ability to export code as Python or TypeScript packages. The architecture allows for deep customization, enabling users to swap models, add tools, or replace components while maintaining consistency between local prototypes and production environments.
