The Strands Agents team at AWS released Strands harness today under the Apache 2.0 license. This open-source agent framework cuts token costs by 28% across six industry benchmarks while matching leading accuracy numbers. Developers often face steep hurdles when moving autonomous systems from prototype stages into production.

Many teams build a working prototype inside environments like Claude Code or Codex. Yet these agent workflows often stall or break when engineers build custom loops from scratch. The Strands harness package bridges that gap by offering a fully assembled agent runtime. The framework supports Python and TypeScript codebases and starts running with a single line of code.

Now teams can run agent workloads on local hardware or deploy them directly to cloud providers. The release tackles token waste, runaway context windows, and execution loops across diverse tasks.

What is Strands harness and How Does It Function?

An agent harness represents the core architecture surrounding a foundation model. This structure includes the execution loop, system tools, context handlers, memory stores, and crash recovery routines. Before this release, the team provided those modular pieces through the Strands Harness SDK. But developers had to stitch the components together manually.

In turn, the new release packages those lower-level primitives into robust defaults. The framework functions as a general agent rather than a dedicated coding tool. By default, calling create_harness() produces an agent ready for broad operational duties.

First, the agent connects directly to modern reasoning models through Amazon Bedrock, Anthropic, OpenAI, and Google. Plus, it routes calls through local runtimes like Ollama or gateways like LiteLLM. Instead of crafting custom tools for each workflow, the harness includes shell access, web retrieval, and file tools out of the box.

How Does Strands harness Handle Model Calls?

The execution loop inside Strands harness treats large tool outputs with special care. When external commands generate heavy text outputs, the system moves the content to disk. Then it caches repeated prompt elements across consecutive turns.

So the system keeps long-term memory intact between separate runs. Users can pause a complex task and resume later using a session identifier. Next, the primary loop can delegate open-ended subtasks to an internal helper agent. It monitors complex steps using an internal checklist. Also, the harness auto-loads dynamic tools from the Agent Skills library whenever it detects them in a repository.

Benchmark Results Across Six Workloads

To prove efficiency, the engineering team executed distributed benchmarks on Amazon EC2. They evaluated the framework with Harbor, an evaluation harness built by the creators of Terminal-Bench. The team averaged results across six distinct suites: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1.

Each benchmark measured average dollar cost per completed task alongside solution accuracy. The benchmark suite tested Strands harness against several prominent agent harnesses. These rivals included Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness.

Still, one nuance stands out in the benchmark data. DeepSeek Harness consumed the fewest tokens overall, running roughly 14% cheaper than the Strands harness suite. But it scored lower on every single benchmark test. Factoring in DeepSeek Harness actually dragged the overall average savings figure down to 28%. Meanwhile, the top performance on the board belonged to Claude Opus 5 running on Strands harness, which reached nearly 85% accuracy.

Why Does Strands harness Cut Costs by 28 Percent?

A direct comparison on Terminal-Bench 2.1 offers a clear view of performance. In this trial, researchers tested Claude Fable 5 across 89 separate runs per harness.

• Strands harness: $56.29 run cost, 69.7% accuracy
• oh-my-pi: $86.83 run cost, 69.7% accuracy
• OpenCode: $73.42 run cost, 66.3% accuracy
• Claude Code: $248.05 run cost, 61.8% accuracy
• DeepSeek Harness: $40.30 run cost, 59.5% accuracy

Against Claude Code, Strands harness slashed operational costs by 77% while improving accuracy by 7.9 percentage points. Even so, oh-my-pi achieved identical 69.7% accuracy, but at a 54% higher financial cost.

While DeepSeek Harness cost less, it trailed by 10.2 percentage points in task completion. Researchers noted that open harnesses hold a distinct cost advantage over proprietary agent wrappers.

Context Management Rules Driving Efficiency

The team points to intelligent context management as the main driver of savings. Without strict controls, automated loops fill context windows quickly. That inflation drives up token bills and degrades reasoning quality over multi-step workflows. Strands harness applies three automated rules to preserve model attention:

• The runtime truncates any tool output that exceeds approximately 1,500 tokens.
• Automatic compaction triggers whenever context consumption crosses 85% of capacity.
• Context recovery procedures run directly inside the main loop if an overflow occurs.

These automated rules align with recent independent findings from the HarnessTax study. That research compared Claude Code, Codex CLI, and Pi across seven distinct model families. The researchers discovered that harness choice rarely altered baseline model capability. Rather, identical models achieved similar completion rates while burning up to five times more budget under unoptimized harnesses.

Thus, controlling token bloat delivers major savings without compromising intelligence. The team confirmed that a detailed technical paper exploring these benchmark findings will arrive soon.

How Can Developers Deploy Strands harness Today?

Getting started with Strands harness takes very little setup. Developers can install the core Python library using pip install strands-harness. Alternatively, TypeScript developers can run npm install @strands-agents/harness.

Once installed, a developer can initialize the runtime with minimal code:

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")

Developers can select any model name or route requests locally through an Ollama endpoint. For rapid agent prototyping, the team also introduced the Strands command line interface via npm install @strands-agents/strands-cli.

The interactive CLI lets builders configure complex agent pipelines using natural language prompts. In a public test, engineers instructed an agent to attach the Playwright MCP server and test video latency on a remote web page. Running /export generated clean Python or TypeScript project files containing the complete configuration.

Strands engineer Gautam Sirdeshmukh also utilized the CLI to build a desktop application that triggers agent runs on remote clusters. Customization remains flexible throughout the stack. Developers can override prompt caches, plug in specialized tools, or rewrite core execution routines down to the underlying SDK.

Because the harness functions as an open library, developers run the exact same logic on laptops as in production environments. A bundled skills module helps agents create deployment configurations for AWS, GCP, Azure, Cloudflare, and Modal.

As teams scale autonomous systems, cutting-edge agent architecture continues to redefine how enterprises manage computational budgets.