- Published on
AWS Strands Agents Team Releases Strands Harness: An Open-Source Agent Harness With 28% Lower Token Cost at Comparabl...
- Authors
- Name
- DeepHorus
The Strands Agents team at AWS has released Strands Harness, an open-source, general-purpose agent framework designed to bridge the gap between prototype and production. Many developers find their agent concepts work well in environments like Claude Code but falter during rebuilding with custom loops. Strands Harness addresses this by providing a fully assembled system that runs locally or deploys to major cloud providers, including AWS, GCP, and Azure.
The harness is a complete system encompassing the model's operating loop, tool management, context handling, memory, and recovery. Unlike specialized coding agents, Strands Harness is built as a general-purpose tool. It ships for Python and TypeScript under the Apache 2.0 license and can be started with a single line of code. The package includes default tools for shell, file operations, and web access.
Key features out of the box include support for models from various providers via Amazon Bedrock, Anthropic, OpenAI, and others. It handles large tool outputs by offloading them to files, caches request components, maintains long-term memory across sessions, and delegates subtasks to a helper agent. The team reports it delivers a 28% lower cost than competing harnesses running the same Claude or GPT models.
This efficiency claim is based on distributed benchmarking on Amazon EC2 using the Harbor framework. The cost saving was measured as an average across six diverse benchmarks, including ALFWorld and Terminal-Bench 2.1, while maintaining near-equal accuracy. On the Terminal-Bench 2.1 test with Claude Fable 5, Strands Harness scored 7.9 points higher than Claude Code at 77% lower cost.
The team attributes the efficiency gains primarily to sophisticated default context management. Three core rules drive this: truncating large tool results, summarizing context when usage exceeds 85%, and recovering from context overflow within the loop. This approach aligns with independent research suggesting harness choice has minimal impact on success rates but can drastically affect cost.
Developers can install the harness via pip for Python or npm for TypeScript. A command-line interface allows prototyping agents in plain English. The release provides a new open-source option aiming to standardize and reduce the cost of deploying capable AI agents across various tasks and cloud environments.
