Strands Harness is a new open-source package from Strands Agents for turning a language model into a general-purpose agent. Instead of wiring the surrounding software by hand, users can start with defaults for the system prompt, tools, context management, sessions, memory, and a built-in helper agent. The package is available for Python and TypeScript and can target hosted models or a local Ollama model.
The Strands Agents launch graphic identifies the package discussed in the article and its focus on turning language models into configurable agents.
The central cost claim comes from Strands’ own benchmarking. The company says that, across six benchmarks using the same Claude or GPT models, its harness cost 28% less than other agent harnesses while maintaining equal or better accuracy. It also reports that Strands Harness cost 77% less than Claude Code with Fable 5 and scored higher on Terminal Bench 2.1. Those are publisher-reported results, not independent validation.
The documentation explains how the harness is designed to reduce repeated work. It can set aside bulky tool results, cache reused portions of requests, retain long-term memory across runs, resume a session by ID, delegate open-ended subtasks, and track multi-step work with a checklist. Its tuned prompt asks the agent to explore before acting, confirm irreversible actions, and verify its work before finishing.
The main limitation is that the supplied launch post points readers to a follow-up paper rather than including the benchmark details needed to assess significance, workload selection, or statistical variation. The supplied documentation also describes the defaults and customization options, but it does not independently test the 28% or 77% cost figures. Readers should therefore treat the numbers as Strands’ reported results and verify them against the promised paper or their own workloads.
The practical distinction is between a ready-made harness and the lower-level Strands Harness SDK. Harness gives builders a preconfigured starting point with replaceable components; the SDK exposes the loop and primitives for constructing an agent piece by piece. That makes Harness most relevant to teams that want a customizable default setup without committing to every underlying design choice.
The documentation graphic supports the article’s discussion of the Harness package, its replaceable components, and the distinction from lower-level SDK primitives.