Amazon Web Services has introduced Strands Harness, an open-source AI agent designed to make it easier for developers to move agentic applications from local prototypes into production environments.
The project addresses a growing problem in AI development. Tools such as Claude Code and Codex make it relatively easy to build and experiment with agents locally, but deploying those systems at scale requires developers to solve additional problems involving context management, memory, tools, execution and infrastructure.
AWS wants Strands Harness to provide much of that foundation out of the box. Despite being developed by AWS, the framework is not restricted to its cloud. It can run locally or on infrastructure from providers including Google Cloud, Microsoft Azure, Cloudflare and others. (SiliconANGLE)
Model and Cloud AgnosticStrands Harness supports models from multiple providers, including Anthropic, OpenAI, Google and Amazon Bedrock. Developers can also connect locally hosted models through technologies such as Ollama. (GitHub)
That portability is a central part of the project’s design. Developers can change the underlying model or deployment infrastructure without rebuilding the entire agent architecture.
The framework includes common agent capabilities such as file reading and writing, editing, shell execution and web search. It also supports the Model Context Protocol (MCP), allowing agents to connect with external tools and services. (SiliconANGLE)
Managing Context More EfficientlyContext management is another major focus.
Long-running agents can quickly fill their context windows with command outputs, documents and previous interactions. That increases token consumption and can eventually reduce the amount of useful information available to the model.
Strands Harness can offload large tool results into files rather than keeping everything directly inside the active context. It can also cache portions of requests that are repeatedly reused. (SiliconANGLE)
The agent maintains persistent sessions as well, allowing developers to resume previous conversations using session identifiers rather than starting from scratch.
A built-in helper agent can handle open-ended subtasks, while developers can extend agents through reusable Agent Skills and external MCP servers. (SiliconANGLE)
AWS Claims Significant Efficiency ImprovementsAWS says its benchmarks show Strands Harness can operate 26% more efficiently than agents built using competing frameworks when using the same underlying model.
In another test involving Anthropic’s Fable 5 model, AWS says Strands Harness achieved a higher score on Terminal Bench 2.1 while costing 77% less than Claude Code for the same tasks. These figures come from AWS’s own testing and will need broader independent benchmarking to establish how representative they are across real workloads. (SiliconANGLE)
The emphasis on token efficiency is significant because agentic applications can generate extremely large numbers of model calls. Small improvements in context management can translate into substantial cost reductions when thousands of agent sessions are running simultaneously.
From Natural Language to an Agent ApplicationOne of the first applications built with the technology is the Strands CLI.
It allows developers — and potentially users with limited programming experience — to prototype agents using natural-language instructions. Users choose a model, provide prompts and tools, and interact with the resulting agent through the command line.
Once the prototype behaves as intended, an/exportcommand can generate the underlying harness as Python or TypeScript code. (SiliconANGLE)
This provides a path from experimentation to a conventional software project without requiring developers to manually recreate the prototype.
Another Step Toward Portable AI AgentsStrands Harness reflects a broader evolution in agent development.
The challenge is increasingly moving beyond simply connecting an LLM to several tools. Production agents require memory, context management, execution controls, observability, guardrails, multi-agent orchestration and reliable infrastructure.
The wider Strands ecosystem already provides multi-agent patterns, tracing, evaluations, guardrails and support for deployments ranging from Docker and Kubernetes to AWS Lambda, Fargate and EKS. (GitHub)
AWS’s decision to make the framework model- and cloud-agnostic is particularly notable. Instead of making the agent itself dependent on AWS infrastructure, Strands attempts to make the agent harness portable.
That could become increasingly important as organizations experiment with different models and infrastructure providers. Rather than rebuilding their agent every time the underlying model or cloud changes, developers could maintain the same agent architecture and swap the components underneath it.
Strands Harness is open source under the Apache 2.0 license and supports both Python and TypeScript. (GitHub)