AWS releases Strands Harness, an open-source agent harness that cuts token costs by 28%
The AWS Strands Agents team has released Strands Harness, an open-source agent harness under the Apache 2.0 license designed to close the gap between a prototype that works in Claude Code or Codex and a version that runs in an autonomous loop. The package is available for Python and TypeScript, installs in a single line, and supports local execution or cloud deployment, including a skills file that generates configuration for AWS, GCP, Azure, Cloudflare and Modal.
What a harness is and why it matters
A harness is the wrapper around the model: the loop, tools, context management, memory and recovery mechanisms. Until now the Strands SDK exposed these building blocks separately; Strands Harness packages them into working defaults. The `create_harness` function returns a general-purpose agent — not a specialised coding agent — that runs on current reasoning models via Bedrock, Anthropic, OpenAI, Google, Ollama or LiteLLM. It ships with shell, file (read, write, edit) and web tools instead of a dedicated tool per task. Heavy tool results are offloaded to files, repeated prompt segments are cached, long-term memory persists across runs, and a session can be resumed by ID. A built-in helper agent accepts open-ended subtasks, and a checklist tracks multi-step work.
The benchmarks and the 28%, with an important caveat
The team ran distributed benchmarking on EC2 using Harbor, the evaluation framework from the creators of Terminal-Bench. The score is an average across six benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench and Terminal-Bench 2.1. Cost was measured in average dollars per task. Competitors in the chart: Claude Code, Codex, oh-my-pi, OpenCode and DeepSeek Harness.
The standout figure — 28% token savings at comparable accuracy — comes with a caveat. DeepSeek Harness was the most token-efficient, roughly 14% cheaper than Strands Harness, but scored lower on every benchmark. The footnote on the chart notes that its inclusion pulled the overall savings figure down to 28%. The highest point on the chart is Claude Opus 5 on Strands Harness, at roughly 85% success.
Head-to-head on Terminal-Bench 2.1
The cleanest comparison uses Claude Fable 5 on Terminal-Bench 2.1 with 89 runs per harness. Against Claude Code, Strands Harness cost 77% less and scored 7.9 points higher. oh-my-pi reached the same accuracy (69.7) at a 54% higher cost. DeepSeek Harness was cheaper still but trailed by 10.2 points. The team notes that two other open-source harnesses also showed strong cost-accuracy performance versus Claude Code.
What drives the efficiency: context management in three rules
Strands Harness ships with defaults for prompt caching and context management. The researchers say context management is the primary driver of both token efficiency and accuracy. Three rules do the work: tool results above roughly 1,500 tokens are truncated; compaction triggers when context usage exceeds 85%; and context recovery runs inside the loop if the window overflows.
This aligns with recent independent research. The HarnessTax study compared Claude Code, Codex CLI and Pi across seven models and found that harness choice barely moved success rates, while the same model achieved similar success at up to a 5× cost difference. Strands researchers say a follow-up paper on their benchmarks is in progress.
Getting started
Install with `pip install strands-harness` or `npm install @strands-agents/harness`. Choose 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")
```
The Strands CLI (npm install...) enables quick launch from the terminal.