The Pulse
Strands harness claims 28% lower token costs across six benchmarks
Strands Agents has released Strands harness, an open-source general-purpose agent that runs locally or on cloud providers and reports lower token use with nearly equal benchmark scores.
AI.info Team ·
Strands Agents released Strands harness on September 21, 2026, as an open-source package for running a general-purpose AI agent locally or deploying it to a cloud provider. The project is available under the Apache 2.0 license and is designed to provide a preassembled set of capabilities without requiring developers to build an agent loop from scratch.
In the announcement, Arron Bailiss, Tim Moreton and Albert Zhao, listed as the post’s authors, wrote:
Today we’re releasing Strands harness: a fully assembled state-of-the-art agent harness you can easily run locally or deploy to your favorite provider.
The authors say Strands harness costs 28% less when the same Claude or GPT models are used across six benchmarks. They also report that it showed better token efficiency and nearly equal benchmark scores compared with Claude Code, Codex and other popular harnesses. The claim is therefore tied to the specific benchmark comparison rather than presented as a general reduction for every use case.
A ready-made agent for general-purpose tasks
Strands harness is built for general-purpose work rather than coding alone. Developers can call create_harness() in Python or createHarness() in TypeScript, select a model and give the resulting agent a task.
The examples in the announcement ask the agent to research several databases, compare pricing and limits, and write the results to a file. The package includes shell, file and web tools, along with long-term memory and the ability to resume an earlier conversation using a session ID. It can also delegate open-ended subtasks to a helper agent, track multi-step work with a checklist and load skills when they are available.
Strands harness supports models through Amazon Bedrock, Anthropic, OpenAI and Google, as well as local models through Ollama and routing through LiteLLM. The post describes deployments on providers that support Linux containers, including Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS and Amazon Bedrock AgentCore.
Developers can install the package with pip install strands-harness for Python or npm install @strands-agents/harness for TypeScript.
Context management and token use
The announcement attributes much of the reported token efficiency to default context-management features. Tool results larger than about 1,500 tokens are truncated, summarization begins when the context window passes 85%, and context recovery runs within the loop if an overflow occurs. Prompt caching stores reused parts of requests to reduce repeated work.
The testing setup used distributed benchmarking on Amazon EC2 with Harbor. The post says Strands harness produced nearly equal scores to Claude Code, Codex and other popular harnesses while using fewer tokens. It also says a follow-up paper from the researchers will provide more information about the benchmarks.
Fable 5 comparison
One comparison in the release focuses on Fable 5. In that test, Strands harness used 77% less cost than Claude Code and scored higher on Terminal Bench 2.1.
The post also identifies Deepseek Harness as the most token-efficient system overall, while saying it typically reported the lowest accuracy scores. These results come from the Strands Agents team’s published comparison; the announcement does not provide the complete methodology, model-version details or a full table of benchmark scores.
From prototype to deployable code
Strands Agents also provides a command-line interface for prototyping. Developers can describe an agent in plain English, select a model provider and add prompts and tools. The CLI can export the resulting project in Python or TypeScript, allowing developers to continue working with the generated code.
The harness remains customizable after setup. Developers can override defaults, switch models, add tools or replace components down to the Strands Harness SDK, which is intended for building an agent harness from the ground up.
The concrete release is a Python and TypeScript package with preselected defaults, support for multiple model providers, built-in tools, context management and published cost comparisons. Developers can install it from PyPI or npm and begin with a single factory call.