Amazon Web Services shipped a new open source framework this week that hands developers a fully assembled AI agent, minus the one component AWS refuses to bake in: the model itself. The Strands harness, released Monday according to The Letter Two, plugs into Amazon Bedrock, Anthropic, OpenAI, Google, or a local Ollama instance, letting whoever installs it swap the underlying model without rebuilding the agent around it.
That modularity is the pitch, but the cost claims are what AWS is leading with. The company says Strands harness runs 26 percent cheaper than competing setups “using the same model across seven benchmarks,” and in one comparison using Fable 5 as the model, AWS reported a 77 percent cost advantage over Claude Code on Terminal-Bench 2.1, though the benchmark’s own maintainers have since moved past that release to version 4.0. Those are AWS’s own numbers, measured on AWS’s chosen benchmark version.
The harness extends Strands Agents, the open source SDK AWS introduced in May 2025 for building agents in a few lines of code. What is new is everything wrapped around the model: long-term memory that resumes a prior session by ID, a helper agent that takes on open-ended subtasks, and a self-maintained checklist for tracking multi-step work. Marc Brooker holds the title of distinguished engineer and vice president at AWS, and in an email to The AI Economy he framed the harder problem as tuning those pieces, not assembling them: “An SDK gives you the building blocks, but you still need to decide how to manage context, integrate tools, and tune the system prompt.”
Brooker described the biggest internal obstacle as context management: long tasks generating thousands of lines of tool output filled the model’s context window and made costs compound. AWS’s fix keeps what’s still relevant, summarizes older turns, and pushes bulky tool output into storage that the agent recalls through a brief lookup instead of reloading everything. Brooker said the payoff was keeping the agent from losing track of its own task while its costs stayed predictable, and he called that “one of the hardest engineering problems” his team had faced.
AWS is not the only company selling a batteries-included agent. Google has Antigravity, Microsoft has its own Agent Framework Harness, and AWS already had AgentCore before this release. The pattern across all four is the same: eliminate the setup work so developers spend time on what the agent does rather than how it is wired together. Strands harness is aimed specifically at long-running operations, DevOps and back-office automation, and in-product conversational features, not coding. “It’s not a coding assistant,” Brooker said. “That’s a different category with great products already in it.”
One comparison AWS did not make in its own favor: DeepSeek’s harness beat Strands harness on token efficiency in the same evaluation where Strands claimed a cost edge over Claude Code. AWS’s announcement does not explain that gap, and it leaves open whether the harness’s cost advantage holds once a team optimizes for tokens rather than dollars.
Teams evaluating an agent harness this quarter should ask any vendor, AWS included, to show benchmark results run on the current version of Terminal-Bench rather than the one AWS cited, since the maintainers have already superseded it.
Reported by Ken Yeung for The Letter Two on September 21, 2026.
