AI
AWS releases open-source AI agent framework priced 45% below Claude Code
Amazon open-sources Strands Harness, a general-purpose agent for developers that bundles tooling, context management, and cost optimization—runnable locally or on any cloud.
The New Stack
Amazon Web Services has unveiled Strands Harness, an open-source AI agent framework that packages pre-configured tools and infrastructure for developers to build and deploy intelligent agents without vendor lock-in.
Built atop Strands Agents, an open-source Python SDK AWS introduced in May 2025, Strands Harness abstracts away low-level decisions by bundling file access, shell commands, web browsing, context memory, session persistence, and prompt caching. According to Marc Brooker, VP and distinguished engineer at AWS, the layer sits between raw SDK building blocks and production deployment, letting developers "focus on their application's task and domain expertise, while customizing the components that need different behavior."
How it works#
Out of the box, Strands Harness provides a working agent capable of multi-step reasoning and tool delegation. Developers can configure which model powers it (Anthropic, OpenAI, Google, or local via Ollama), adjust instructions and capabilities per agent, and chain multiple agents together. The framework's core logic—looping, tool execution, memory, and delegation—runs locally by default; model inference can route through Amazon Bedrock but can be swapped with one line of code.
A command-line interface enables interactive prototyping, letting developers pick a model, add prompts and tools, then export the result as Python or TypeScript. AWS also bundles an Agent Skill to help coding agents understand Strands Harness internals and generate deployment configs for AWS, GCP, Azure, Cloudflare, and Modal.
Cost and performance gains#
AWS benchmarked Strands Harness against Claude Code and Codex using six standard test suites (ALFWorld, ContextBench, GAIA, WebShop, τ³-bench, and Terminal-Bench 2.1), reporting 45% lower cost while maintaining comparable accuracy. That margin narrows to 28% when including DeepSeek Harness, which AWS notes runs 14% cheaper.
On Terminal Bench 2.1 specifically, Strands Harness with Fable 5 model cost $56.29 across 89 trials (score: 69.7) versus Claude Code's $248.05 (61.8). AWS attributes the savings to context-management defaults: truncating large tool outputs, compacting context windows when needed, and graceful overflow recovery—decisions that would otherwise fall to individual developers.
