AWS Strands Brokers Staff Releases Strands Harness: An Open-Supply Agent Harness With 28% Decrease Token Price at Comparable Accuracy

Many builders discover that an agent concept works inside Claude Code or Codex, then struggles as soon as they rebuild it with their very own loop. The Strands Agents group at AWS is concentrating on that hole with Strands harness, a completely assembled, general-purpose agent harness. It runs domestically or deploys to a cloud supplier, ships for Python and TypeScript below Apache 2.0, and begins with one line of code. The group stories 28% decrease price than different harnesses operating the identical Claude or GPT fashions throughout 6 benchmarks, with near-equal accuracy.

Is it deployable? Sure. It runs domestically, and a bundled abilities file helps your coding agent generate deployment config for AWS, GCP, Azure, Cloudflare, and Modal.

What’s Strands Harness

A harness is the system across the mannequin: the loop, instruments, context dealing with, reminiscence, and restoration. Strands already uncovered these constructing blocks by the Strands Harness SDK. Strands harness packages them into working defaults. It’s constructed as a general-purpose agent, not a coding agent.

Out of the field, create_harness() returns an agent that:

• Runs on a present reasoning mannequin by Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM.
• Ships shell, file (learn, write, edit), and internet instruments, as an alternative of a bespoke device per activity.
• Offloads cumbersome device outcomes to recordsdata and caches reused elements of every request.
• Retains long-term reminiscence throughout runs and resumes a dialog from a session ID.
• Delegates open-ended subtasks to a built-in helper agent and tracks multi-step work with a guidelines.
• Masses Agent Skills when it finds them.

Benchmark Setup and the 28% Determine

The Strands Agents group ran distributed benchmarking on Amazon EC2 with Harbor, the analysis framework from the Terminal-Bench creators. The rating is the typical throughout 6 benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1. Price is the typical {dollars} per activity. Rivals on the chart are Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness.

One necessary factor to notice. DeepSeek Harness was essentially the most token-efficient harness total, operating about 14% cheaper than Strands harness. It additionally scored decrease on each benchmark. The chart footnote states that together with it introduced the general financial savings determine down to twenty-eight%. The best-scoring level on the chart is Claude Opus 5 on Strands harness, close to 85%.

Terminal-Bench 2.1: Identical Mannequin, 5 Harnesses

The clearest head-to-head makes use of Claude Fable 5 on Terminal-Bench 2.1, with 89 trials per harness.

In opposition to Claude Code, Strands harness price 77% much less and scored 7.9 factors increased. Oh-my-pi matched its 69.7 accuracy at 54% increased price. DeepSeek Harness was cheaper nonetheless, however trailed by 10.2 factors. The group additionally famous that 2 different open-source harnesses carried out effectively on price and accuracy towards Claude Code.

What Drives the Effectivity

Strands harness ships defaults for immediate caching and context administration. The group says context administration largely drove each token effectivity and accuracy. 3 guidelines do the work:

• Instrument outcomes over about 1,500 tokens get truncated.
• Summarization (compaction) triggers when context utilization passes 85%.
• Context restoration runs contained in the loop if the window overflows.

This matches current impartial analysis. The HarnessTax study in contrast Claude Code, Codex CLI, and Pi throughout 7 fashions. It discovered harness selection barely moved success charges, whereas the identical mannequin reached comparable success at as much as 5x the price. The Strands researchers say a follow-up paper on their benchmarks is coming.

Getting Began

Set up with pip set up strands-harness or npm set up @strands-agents/harness. Decide a mannequin by identify, or level the harness at a neighborhood Ollama mannequin:

from strands_harness import create_harness

agent = create_harness(mannequin="litellm/openai/gpt-5.6-sol")
agent("Research the top three vector databases and compare their pricing")

The Strands CLI (npm set up @strands-agents/strands-cli) allows you to prototype an agent in plain English. Within the group’s demo, the agent was requested so as to add the Playwright MCP server and measure video load latency on a weblog publish. Working /export then produced the harness code, with the Playwright MCP included, as a Python or TypeScript zip.

The CLI itself is constructed on Strands harness. Strands engineer Gautam Sirdeshmukh additionally used it to construct a desktop app that begins Strands harness runs remotely.

Customization goes deep. You possibly can override any default, swap fashions, add instruments, or change parts right down to the Strands Harness SDK. As a result of the harness is a library dependency, the agent prototyped on a laptop computer is identical one embedded in manufacturing.

Key Takeaways

• Strands harness packages AWS’s Strands primitives right into a general-purpose, Apache 2.0 agent.
• It stories 28% decrease price than rival harnesses throughout 6 benchmarks at comparable accuracy.
• With Fable 5 on Terminal-Bench 2.1, it price 77% lower than Claude Code and scored increased.
• Context defaults drive the beneficial properties: 1,500-token truncation, 85% compaction, in-loop restoration.
• One create_harness() name targets Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM.

Take a look at the Technical details, GitHub repo, PyPI package, and Strands Agents docs.