Amazon Web Services (AWS) announced on September 29, 2026, the availability of the open-source Strands Harness tool, a layer for composing AI agents capable of performing multiple tasks using large language models from different providers, rather than being tied to a single model or to the AWS environment alone.
The tool is based on the Strands Harness SDK that AWS published in August 2026. It supports models running through Amazon Bedrock, in addition to Anthropic Claude, OpenAI GPT, Google Gemini, Ollama, and LiteLLM. According to the original article, an agent can be created with just a few lines of code, and the model can later be changed by modifying its configuration value.
What does the tool provide?
An agent created through Strands Harness comes with built-in capabilities including executing shell commands, reading and writing files, and searching the web. The tool also includes mechanisms for prompt caching and conversational context management, functions that directly affect token consumption and the continuity of task execution.
According to the stated default settings, tool results exceeding approximately 1,500 tokens are truncated, and context compression through summarization begins when context-window usage exceeds 85%. If the capacity is exceeded, the tool attempts to restore the context within the execution loop. Session memory can also be retained and a previous conversation resumed using the session ID, with support for what AWS calls “skills.”
Model and deployment flexibility
The main practical point is separating the agent layer from the language model and the deployment destination. After the agent is built, the model being used can be switched, and the agent can be deployed in any environment where Linux containers are available, not only in AWS services. The open-source nature makes it possible to add custom functions and modifications, but the article does not provide details about operating requirements or the governance and security mechanisms needed for each environment.
AWS also provides a command-line interface called Strands CLI that allows the model, tools, extensions, and agent settings to be selected interactively. After building, the agent can be exported as a TypeScript or Python source file.
What changes in practice?
Strands Harness makes building agents closer to a reusable assembly process: the developer specifies the model, tools, and tasks, then can change the model or deployment environment without rebuilding the entire agent. This may benefit teams that want to test multiple models or avoid tying their applications to a single provider, while results quality and cost remain linked to the selected model and the nature of the tools and tasks.
AWS also published a benchmark comparison in which it said that an agent built through Strands Harness using Claude achieved, in a test based on Fable 5, a score of 69.7 at a cost of $56.29, presenting it as less expensive and more capable than Claude Code in the comparison shown. However, these figures come from AWS itself and are not sufficient on their own to assess performance outside the test scenario.
This means the announcement is not limited to a simplified interface for creating an agent; it presents an open-source framework for managing tools, context, memory, and deployment. At the same time, granting the agent permission to execute commands and access files and the web makes permission controls, governance, and audit trails essential operational questions, and the source material did not include enough details to answer them.