Amazon Web Services has released Strands Harness, a new open source AI agent that runs 45% cheaper than competing tools while matching their accuracy, according to the company's internal testing. The general-purpose agent, announced through The New Stack, gives developers a preconfigured foundation they can operate locally or push to any cloud provider. Built atop the Strands Agents SDK that AWS launched in May 2025, Strands Harness packages together the machinery an agent needs for extended tasks, with AWS providing its own defaults for how those components interact.
Out of the package, Strands Harness delivers a functioning agent equipped with file, shell, and web tools, plus integrated handling for context, memory, ongoing sessions, prompt caching, and delegation to other agents. Developers install it as a Python or TypeScript package through pip install strands-harness or npm install @strands-agents/harness. AWS averaged Strands Harness scores across six benchmarks—ALFWorld, ContextBench, GAIA, WebShop, τ³-bench, and Terminal-Bench 2.1—then compared that with average cost per task across identical tests. Against Claude Code and Codex specifically, the company says Strands Harness ran 45% cheaper with broadly similar accuracy. That advantage shrinks to 28% when DeepSeek Harness, which AWS says operated roughly 14% cheaper than Strands Harness on matched runs, joins the broader comparison. On Terminal Bench 2.1 alone, Strands Harness running Fable 5 cost $56.29 across 89 trials versus Claude Code's $248.05, a 77% reduction, while scoring 69.7 against 61.8.
Marc Brooker, VP and distinguished engineer at AWS, tells The New Stack that developers still need to "decide how to manage context, persist conversations, integrate tools, and guide the agent's behavior." According to Brooker, an SDK like Strands Harness supplies the building blocks, but developers retain control over instructions, model selection, tool availability, and whether agents can hand work to another. The report notes that most of Strands Harness operates independently of AWS infrastructure—the agent loop, tools, context management, session handling, and delegation ship in the open source release and run on the machine hosting the agent by default. The lone exception is the call to the underlying model, which AWS routes through Amazon Bedrock, its managed service for accessing foundation models, though Brooker says "this is easily overrided to use a different model provider with one line." Strands Harness can instead tap Anthropic, OpenAI, Google, or a locally running model through Ollama.
AWS attributes the cost advantage largely to its context-management defaults: Strands Harness truncates especially large tool outputs, compacts context once the available window crosses a set threshold, and tries to recover within the agent loop if context overflows. Brooker says work on Strands Harness can feed improvements into Amazon Bedrock AgentCore, the company's managed service for deploying and operating agents, and vice versa, giving AWS a path for technology developed in the open source project to inform its managed offering. The arrangement hands AWS a clear commercial route: developers adopt Strands Harness freely, while AgentCore becomes a natural destination for teams that eventually want AWS to manage the infrastructure around it. Brooker stresses that Strands Harness can deploy independently of AgentCore, outside AWS altogether, with all deployment paths open for developers to choose. For AWS, packaging and tuning functions like context management makes more sense than requiring every developer to puzzle through those decisions independently with the SDK, Brooker says, noting that "getting a prototype working is one step; evaluating how those choices affect performance and cost is another." By offering a ready-made agent that developers can tailor or replace piece by piece, AWS positions itself at the center of a market where cost efficiency and model flexibility increasingly determine which tools gain traction. Organizations weighing open source control against managed convenience will likely face this exact trade-off repeatedly as agent architectures mature.