AWS Strands Agents Team Introduces Strands Harness: An Open-Source Agent Harness That Cuts Token Costs by 28% While Maintaining Comparable Accuracy

If you've ever built an agent inside Claude Code or Codex, only to watch it fall apart when you rebuilt it with your own loop, a new release from AWS is worth your attention. The Strands Agents team has launched Strands harness, a fully assembled, general-purpose agent harness designed to close that gap between prototype and production. It runs locally or deploys to any major cloud, supports Python and TypeScript under the Apache 2.0 license, and gets you started with a single line of code. According to the team, it delivers 28% lower cost than competing harnesses running the same Claude or GPT models across 6 benchmarks, with accuracy that stays essentially the same.

Can you actually deploy it? Yes. Beyond running on your own machine, the package includes a skills file that helps your coding agent produce deployment configurations for AWS, GCP, Azure, Cloudflare, and Modal.

What Exactly Is Strands Harness?

A harness is everything wrapped around the model itself: the execution loop, tool calls, context management, memory, and error recovery. The Strands Agents SDK already offered these pieces individually. The new harness bundles them into sensible, working defaults — and importantly, it's built as a general-purpose agent, not a coding-specific one.

Out of the box, calling create_harness() gives you an agent that:

• Runs on a current reasoning model through Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM.
• Comes with built-in shell, file (read, write, edit), and web tools, rather than requiring a custom tool for every task.
• Moves large tool outputs to files and caches the parts of each request that get reused.
• Maintains long-term memory between runs and can pick up a conversation again from a session ID.
• Hands open-ended subtasks to a built-in helper agent and follows multi-step work using a checklist.
• Automatically loads Agent Skills whenever it discovers them.

Where the 28% Figure Comes From

The team ran their benchmarks in a distributed setup on Amazon EC2 using Harbor, an evaluation framework built by the creators of Terminal-Bench. The headline number is an average across 6 benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1. Cost was measured as the average dollars spent per task. The competitors shown in their charts include Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness.

One caveat deserves attention. DeepSeek Harness was actually the most token-efficient option overall, running roughly 14% cheaper than Strands harness. However, it scored lower on every single benchmark. A chart footnote explains that including DeepSeek Harness in the comparison brought the overall savings figure down to 28%. The best-performing point on the chart is Claude Opus 5 running on Strands harness, at close to 85%.

Terminal-Bench 2.1: One Model, Five Harnesses

The most direct comparison uses Claude Fable 5 on Terminal-Bench 2.1, with 89 trials per harness:

| Harness | Run cost | Accuracy |
|---|---|---|
| Strands harness | $56.29 | 69.7 |
| Oh-my-pi | $86.83 | 69.7 |
| OpenCode | $73.42 | 66.3 |
| Claude Code | $248.05 | 61.8 |
| DeepSeek Harness | $40.30 | 59.5 |

Compared with Claude Code, Strands harness spent 77% less and scored 7.9 points higher. Oh-my-pi matched its 69.7 accuracy, but at a cost 54% higher. DeepSeek Harness was the cheapest of all, yet trailed Strands by 10.2 accuracy points. The team also noted that 2 additional open-source harnesses beat Claude Code on both cost and accuracy.

What Makes It So Efficient

Strands harness comes with built-in defaults for prompt caching and context management, and the team credits context management for most of the gains in both token efficiency and accuracy. Three rules do the heavy lifting:

1. Tool results longer than about 1,500 tokens get truncated.
2. Summarization (compaction) kicks in once context usage passes 85%.
3. Context recovery runs inside the loop if the window overflows.

These defaults line up with recent independent findings. The HarnessTax study compared Claude Code, Codex CLI, and Pi across 7 models and discovered that the choice of harness barely affected success rates — but the same model could reach similar results at up to 5x the cost difference. The Strands researchers say a follow-up paper covering their benchmarks is on the way.

How to Get Started

Install it with pip install strands-harness or npm install @strands-agents/harness. Then pick a model by name, or point the harness at a local Ollama model:

```python
from strands_harness import create_harness

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

There's also a Strands CLI (npm install @strands-agents/strands-cli) that lets you prototype an agent by describing what you want in plain English. In the team's demo, the agent was asked to add the Playwright MCP server and measure video load latency on a blog post. Running /export then generated the harness code — Playwright MCP included — as a Python or TypeScript zip file.

Notably, the CLI itself runs on Strands harness. Strands engineer Gautam Sirdeshmukh even used it to build a desktop app that launches Strands harness runs remotely.

Customization goes deep. Every default can be overridden, models can be swapped, tools can be added, and components can be replaced all the way down to the Strands Agents SDK. And because the harness is a library dependency, the agent you prototype on your laptop is the same one you embed in production.

Key Takeaways

• Strands harness turns AWS's Strands primitives into a general-purpose, Apache 2.0-licensed agent.
• It reports 28% lower cost than competing harnesses across 6 benchmarks, with comparable accuracy.
• With Fable 5 on Terminal-Bench 2.1, it cost 77% less than Claude Code while scoring higher.
• The context defaults drive the savings: 1,500-token truncation, 85% compaction, and in-loop recovery.
• A single create_harness() call targets Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM.

For more information, check out the technical details, the GitHub repo, the PyPI package, and the Strands Agents docs.