AWS released Strands Harness on September 21, a general-purpose, open-source AI agent that the company says matched rival coding agents on benchmarks while consuming significantly fewer tokens2,4.
Strands Harness builds on the Strands Agents SDK that AWS debuted in May 2025 as an open-source Python toolkit for building AI agents. Marc Brooker, VP and distinguished engineer at AWS, told The New Stack that Strands Harness sits above that SDK, giving developers a preconfigured agent with defaults for context management, tool integration, and long-running task orchestration. AWS later extended Strands to TypeScript and in February created Strands Labs as a home for more experimental projects.
Cost and benchmark claims
AWS says Strands Harness was 45% cheaper than Claude Code and Codex with broadly comparable accuracy. That gap narrows to 28% once DeepSeek Harness is included in the comparison. AWS says DeepSeek Harness ran around 14% cheaper than Strands Harness on matched runs.
The company averaged scores across six benchmarks using the Harbor framework: ALFWorld, ContextBench, GAIA, WebShop, π-bench, and Terminal-Bench 2.1. AWS says Strands Harness consumed 28% fewer tokens across those six tests compared to Claude or GPT models. On Terminal-Bench 2.1, Strands Harness running Fable 5 scored 69.7 versus 61.8 for Claude Code across 89 trials, at a cost of $56.29 versus $248.05. DeepSeek Harness cost $40.30 on the same benchmark but scored 59.5.
The Register noted that AWS "only raced coding agents and marked its own homework" — the benchmarks were run by AWS itself, and Strands Harness is positioned as a general-purpose agent rather than a coding assistant.
What ships in the box
Out of the box, Strands Harness includes file, shell, and web tools, plus built-in handling for context, memory, persistent sessions, prompt caching, and delegation to other agents. The agent defaults to truncating tool results over 1,500 tokens and automatically compacts its context window when it surpasses 85% capacity, attempting recovery if context overflows. It is available as a Python or TypeScript library and also ships with a CLI for prototyping and configuring agents interactively.
"The moment you build your own agent, you're on your own," AWS stated in its announcement. "It's tricky wiring up the right primitives just well enough to match that 'it just worked' feeling".
Provider flexibility
Most of Strands Harness does not depend on AWS infrastructure. The agent loop, tools, context management, session handling, and delegation run on the local machine by default. AWS routes model access through Amazon Bedrock by default, but Strands Harness can use Anthropic, OpenAI, or Google as its model provider, or a locally running model through Ollama. AWS includes an Agent Skill to help coding agents understand Strands Harness and generate deployment configuration for providers including AWS, GCP, Azure, Cloudflare, and Modal.
AgentCore Harness and Strands Harness were built by the same team but live in separate codebases; Strands Harness can be deployed independently of AgentCore and outside of AWS altogether.
ANALYSIS By benchmarking against coding-focused agents while branding Strands Harness as general-purpose, AWS is staking a cost-efficiency claim that may not map cleanly to the broader workloads the agent is designed to handle.