AWS's Strands Agents team has released Strands Harness, an open-source agent framework that dramatically reduces the cost of running AI coding agents while maintaining or improving accuracy. The tool addresses a real pain point for developers: an agent prototype that works in Claude Code or OpenAI Codex often struggles when rebuilt with a custom loop. Strands Harness packages the essential components of a working agent into a single, deployable system that runs locally or on cloud providers like AWS, Google Cloud, Azure, Cloudflare, and Modal.

What Makes Strands Harness More Efficient Than Competitors?

The AWS team benchmarked Strands Harness against rival systems including Claude Code, OpenAI Codex, oh-my-pi, OpenCode, and DeepSeek Harness across six standardized tests: ALFWorld, ContextBench, GAIA, WebShop, tau-squared-bench, and Terminal-Bench 2.1. The results were striking. On Terminal-Bench 2.1 using Claude Fable 5, Strands Harness cost $56.29 per task with 69.7% accuracy, compared to Claude Code's $248.05 per task at 61.8% accuracy. That's a 77% cost reduction with a 7.9 percentage point accuracy improvement.

Across all six benchmarks, Strands Harness reported 28% lower token costs than competing harnesses at comparable accuracy levels. The team noted that while DeepSeek Harness was slightly cheaper overall, it scored lower on every single benchmark, trailing Strands Harness by 10.2 points on Terminal-Bench 2.1.

How Does Context Management Drive These Savings?

The efficiency gains don't come from using better or cheaper models. Instead, they stem from three simple but effective context management rules that reduce unnecessary token consumption. These defaults are built into Strands Harness out of the box:

• Token Truncation: Tool results larger than approximately 1,500 tokens are automatically truncated to prevent bloated context windows.
• Compaction Triggers: When context usage exceeds 85%, the system automatically summarizes and compacts previous interactions to free up space.
• In-Loop Recovery: If the context window overflows despite these measures, the agent recovers gracefully without losing the conversation thread.

This approach aligns with recent independent research. The HarnessTax study compared Claude Code, Codex CLI, and Pi across seven models and found that harness design choices barely affected success rates, yet the same model could reach similar results at up to five times the cost depending on how the harness managed context. The Strands researchers indicated that a follow-up paper detailing their benchmarking methodology is coming.

What Tools and Features Does Strands Harness Include?

Strands Harness ships with practical defaults that cover common agent needs. Out of the box, the framework provides shell access, file operations (read, write, edit), and web browsing tools rather than requiring developers to build custom tools for each task. The agent maintains long-term memory across runs, can resume conversations from a saved session ID, and delegates open-ended subtasks to a built-in helper agent while tracking multi-step work with a checklist.

The system supports multiple model providers. Developers can run agents on current reasoning models through Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM by changing a single parameter. A bundled skills file helps agents automatically generate deployment configuration for major cloud platforms, removing friction from the prototype-to-production pipeline.

How to Get Started With Strands Harness

• Installation: Install the Python package with "pip install strands-harness" or the TypeScript package with "npm install @strands-agents/harness."
• Basic Setup: Create a working agent with just three lines of code: import the library, call create_harness() with your chosen model, and pass a task as a string.
• Prototyping: Use the Strands CLI (npm install @strands-agents/strands-cli) to prototype agents in plain English, then export the generated harness code as a Python or TypeScript zip file.
• Customization: Override any default, swap models, add custom tools, or replace components down to the underlying Strands Harness SDK for deeper customization needs.

The framework is released under the Apache 2.0 open-source license, meaning developers can use, modify, and distribute it freely. Because the harness is a library dependency rather than a proprietary service, the agent prototyped on a laptop is identical to the one deployed in production.

Why Does This Matter for AI Agent Development?

The release addresses a growing frustration in the AI agent community. Many developers report that agents work smoothly in integrated environments like Claude Code or Codex but fail or become prohibitively expensive when moved to custom implementations. Strands Harness bridges that gap by providing a battle-tested, general-purpose agent loop that handles the messy details of context management, memory, and tool integration automatically.

The cost savings are particularly significant for teams running agents at scale. A 77% reduction in per-task expenses could translate to substantial savings for organizations running hundreds or thousands of agent tasks monthly. The accuracy improvements suggest that better context management doesn't just save money; it actually helps agents reason more effectively.

The open-source release also signals a shift in how AI infrastructure is evolving. Rather than locking developers into proprietary platforms, the Strands team is providing a reusable foundation that works across multiple model providers and deployment targets. This approach gives developers flexibility while reducing the engineering overhead of building production-grade agent systems from scratch.