Strands Harness keeps five subsystems local and routes one call to Bedrock

The 45% needs a workload attached before it means anything. The New Stack put the figure in its headline and credited it to AWS ^[1]; the report does not say what task was run or which models were compared ^[18]. Cost in a harness comes from tokens: how much context goes up each turn, how often the cache hits, how many turns the loop takes. Strands Harness ships with built-in handling for context, memory, persistent sessions and prompt caching ^[7]. For 45% to transfer to your workload, the compared runs would need the same task and the same model at the same token price, so that context policy and caching are the only variables left. The harness treats the model provider as a one-line setting ^[10].

Marc Brooker, a VP and distinguished engineer at AWS, said the developer still makes those calls. "An SDK like the Strands Harness SDK gives you the building blocks, but you still need to decide how to manage context, persist conversations, integrate tools, and guide the agent's behavior," he told The New Stack ^[6]. Strands Harness is those decisions taken in advance, with AWS's defaults, sitting above the existing Strands SDK ^[5].

Five of the pieces are in the open source release and, AWS says, run on the machine running the agent by default: the agent loop, tools, context management, session handling and delegation ^[8]. The sixth is the call to the model, and that one goes to Amazon Bedrock ^[9][19]. Brooker said Bedrock is the only out-of-the-box default tied to AWS infrastructure, and that "This is easily overrided to use a different model provider with one line" ^[9][10]. Anthropic, OpenAI, Google and a local model through Ollama are the alternatives named ^[11]. "No features require Bedrock," he said ^[12].

AWS is not selling this as a Claude Code replacement. It calls the harness general purpose, says it takes cues from harnesses such as Claude Code and Codex, and puts the difference in being able to deploy to whichever cloud provider you choose ^[14][15]. The portability is what moves a build-versus-buy calculation, and the CLI makes the artifact concrete: configure the agent interactively, then export it as Python or TypeScript with /export ^[16]. AWS also ships an Agent Skill so coding agents can read the harness itself, including how to add MCP servers or generate deployment configuration ^[17].

One vendor-specific default in an open source release from a cloud provider is restrained. In my view the first thing to test is the export path: whether an agent configured in the CLI and exported to Python behaves the same against a non-Bedrock provider, since Brooker's claim is that context management, sessions, tools and delegation do not change ^[12][16].