Strands harness launches as a fully assembled, general-purpose agent harness runnable locally or on any cloud provider with a single line of Python or TypeScript. Built on the Strands Harness SDK and licensed under Apache 2.0, it supports models from Amazon Bedrock, Anthropic, OpenAI, Google, and Ollama. Benchmarks show Strands harness costs 28% less than Claude Code and Codex using the same models across six tests, while maintaining equal or better accuracy. On Fable 5, it cost 77% less than Claude Code and scored higher on Terminal Bench 2.1. Deepseek Harness was most token-efficient but scored lowest on accuracy.

Default context management drives efficiency: tool results over ~1500 tokens are truncated, summarization triggers at 85% context window, and context recovery runs within the loop on overflow. The agent includes shell, file, and web tools, manages its own context window, caches prompts, keeps long-term memory across runs, delegates subtasks to a helper agent, and tracks work with a checklist.

Strands harness deploys on any Linux container provider including Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS, and Amazon Bedrock Agent Core. The Strands CLI allows prototyping in plain English and exporting code to TypeScript or Python for further iteration.