AWS Releases Strands Harness, a Bring-Your-Own-Model AI Agent at 28% Lower Token Cost

AWS's Strands Harness is a fully assembled agent scaffold designed to run locally or deploy to any provider. Out of the box it can search the web, run commands, edit files, remember previous work, and hand off tasks between steps — the core capabilities of a modern autonomous coding-and-research agent. The differentiator is its bring-your-own-model architecture: rather than locking developers to a single frontier model, Strands lets them plug in whichever model they prefer.

The efficiency claim is the headline. AWS VP of Agentic AI Swami Sivasubramanian announced on X that Strands delivers 'frontier performance at 28% lower token cost than comparable harnesses while matching or beating them on accuracy.' Token cost is the dominant operational expense for agentic workloads — where a single task can consume hundreds of thousands of tokens across many tool-calling turns — so a 28% reduction is directly meaningful to production economics.

Strands rounds out a remarkably dense week of AWS agent infrastructure announcements: AgentCore Runtime V2 (production SLA, ~2s cold starts), continued Bedrock model expansion (Grok 4.6), and now a model-agnostic agent harness. The strategic thread is consistent — AWS is building the full agentic stack as neutral infrastructure, betting that enterprises want model choice, cost control, and governed deployment rather than a single-vendor agent. The harness competes conceptually with offerings from OpenAI, xAI, and open-source frameworks like LangChain, but AWS's angle is the combination of the harness with its runtime, memory, and security layers. The efficiency claim will need independent verification, and the bring-your-own-model flexibility is only as good as the models it hosts — but for AWS-committed enterprises, Strands lowers the barrier to shipping production agents.