AWS Launches Open-Source AI Agent Framework That Slashes Token Consumption by 28%

AWS Strands Tackles the Agent Deployment Challenge

Amazon Web Services has launched Strands Harness, an open-source AI agent framework built to address a persistent challenge developers face when scaling AI applications from local prototypes to production cloud environments1

1

2

2

. Many developers have successfully built AI agent prototypes locally using tools like Anthropic's Claude Code and OpenAI's Codex, but transitioning these setups to scalable infrastructure has proven difficult. AWS Strands eliminates this friction by offering a batteries-included harness that works with just one line of Python or TypeScript, whether deployed locally or across any cloud provider including Google Cloud, Microsoft Azure, Modal, and Cloudflare1

1

2

2

.

Token Efficiency Through Intelligent Context Management

The open-source agentic AI harness demonstrates significant cost advantages through advanced context management techniques. AWS Strands consumed 28% fewer tokens across six benchmark tests compared to agents running Claude or GPT models, with one test showing a 77% cost reduction versus Claude Code while achieving higher scores on the Terminal Bench 2.1 benchmark1

1

2

2

. This efficiency stems from default settings that include automatic prompt caching, truncating tool results over 1,500 tokens, and automatically compacting the context window when it surpasses 85% capacity. The AI agent framework also attempts automatic context recovery in overflow situations, intelligently managing its own context window by offloading tool results to separate files and caching reused portions of requests1

1

2

2

.

Flexible Foundation for Multi-Agent Patterns

Built atop the Strands Harness SDK, this AI agent framework supports multi-agent patterns and integrates with frontier models from Anthropic, OpenAI, Amazon Bedrock, and Google2

2

. Developers preferring local deployments can point the framework to Ollama models hosted on their own machines. The harness ships with essential capabilities out of the box, including read, write, edit, shell, and web search tools that underlying models already know how to use, eliminating the need for bespoke tool design for each agentic task1

1

2

2

. It also maintains long-term memory across runs through session IDs, enabling conversation resumption, and includes a built-in helper agent for subtask delegation using automated checklists2

2

.

Benchmark Performance and Accuracy Questions

AWS tested the framework using the Harbor framework distributed across multiple EC2 nodes, claiming nearly equal benchmark scores versus Claude Code, Codex, and other popular harnesses1

1

. While DeepSeek's harness proved more token efficient, AWS reports it was less accurate on the same tests. However, AWS describes Strands Harness as a general-purpose agent rather than a coding agent, raising questions about the comparison since every competing harness tested—Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness—is marketed as a coding agent1

1

. Amazon Web Services plans to publish a research paper with more detailed benchmark information, which should provide clearer insight into how it performs against other general-purpose agents1

1

.

Rapid Prototyping Through Natural Language Commands

The primary use case centers on rapid prototyping of AI applications, exemplified by AWS teams using Strands Harness to develop the Strands CLI2

2

. This command-line interface enables both developers and non-coders to prototype AI agents using natural language commands. Users simply select the underlying model, add necessary prompts and tools, then use an "/export" command to download the harness code as a Python or TypeScript file. The framework is available now on GitHub and can be installed directly via pip install strands-agents-harness or npm install strands-agents-harness1

1

2

2

. Developers can also upload Agent Skills and integrate external tools such as Model Context Protocol servers, making the platform extensible for specific use cases2

2

. Watch for independent benchmarks that compare this framework against other general-purpose agents to validate AWS's efficiency claims across broader application scenarios.