"Token costs down 28%": AWS unveils 'Strands Harness'

[Financial News] Amazon Web Services (AWS) has unveiled the open-source artificial intelligence (AI) agent 'Strands Harness.' It is an execution environment that bundles together the components needed to run an AI agent.
According to industry sources on the 23rd, developers have valued existing tools such as Claude Code and Codex for providing a smooth experience in which they work effectively without requiring manual intervention. However, building an agent from scratch requires developers to connect multiple basic components individually, making it difficult to replicate that seamless experience. Strands Harness preassembles the process, allowing developers to run an agent with their chosen AI model using just one line of code. It was designed as a general-purpose agent for a wide range of tasks, rather than as a tool limited to specific coding work.
According to AWS's internal tests, Strands Harness reduced token costs by an average of 28% across six benchmarks when using the same Claude or GPT models. Its accuracy remained at or above the level of widely used tools such as Claude Code and Codex. In tests using the Fable 5 model, costs were 77% lower than with Claude Code, while the Terminal-Bench 2.1 score was higher. The tests were conducted in a distributed environment on Amazon EC2 using the Harbor tool, and detailed results will be released in a subsequent paper.
A significant portion of this token efficiency comes from the default context-management settings, which control the amount of information a model can remember and reference at once. Strands Harness automatically summarizes and condenses tool execution results when they exceed approximately 1,500 tokens. It also automatically summarizes the conversation when the amount of information exceeds 85% of its available context. If the information overflows, the system automatically recovers during execution. In addition, prompt caching stores and reuses recurring requests, reducing both processing time and costs.
Strands Harness works with the latest models from Amazon Bedrock, Anthropic, OpenAI, and Google, as well as locally run Ollama models. Developers only need to specify their preferred model by name. The default agent comes equipped with basic tools that models already know how to use, including shell commands, file reading, writing and editing, and web searches. It manages its own information scope by moving extensive execution results to files and caching repeated requests. The system maintains long-term memory across multiple work sessions, and users can resume previous conversations by entering a session ID. Complex tasks can also be delegated to built-in sub-agents, while multi-step tasks are managed through checklists.
Strands Harness can be deployed in any environment that supports Linux containers.
[email protected] Jang Min-kwon Reporter