AWS Strands Harness: The Open-Source Agent Framework Built for Production

Amazon Just Open-Sourced the Missing Piece Between "AI Demo" and "AI in Production"

If you build automation or AI agents for clients, you already know the gap: it takes an afternoon to get an agent working on your laptop, and weeks to get it running reliably somewhere a client can actually rely on. On September 21, 2026, AWS released Strands Harness, a free, open-source agent framework built to close exactly that gap — and it's worth a look whether or not you touch AWS day to day.

What Strands Harness Actually Is

Strands Harness is an open-source SDK (Apache 2.0 licensed) for building AI agents that can run anywhere: on a laptop, inside AWS, or on Google Cloud, Microsoft Azure, Modal, and Cloudflare. It's the productized version of the internal "agent harness" patterns AWS has been using to build its own coding and automation agents, now packaged so any developer can install it with pip install strands-agents-harness or npm install strands-agents-harness.

The framework isn't tied to Amazon's own models. It routes through Amazon Bedrock, Anthropic, OpenAI, Google, local Ollama models, or LiteLLM, and it ships with a set of built-in, general-purpose tools — file read/write/edit, shell commands, and web search — rather than task-specific ones you'd have to build yourself.

Why This Matters for Agencies and Automation Builders

The specific problem Strands Harness targets is the one most agencies run into first: a prototype that works great with Claude Code or a similar coding assistant on a developer's machine doesn't survive contact with a real client environment. Long-running sessions blow through context windows, costs spiral, and there's no clean way to resume a session or hand off a subtask.

Strands Harness bakes in three things that solve this directly:

• Persistent memory across sessions, tracked by session ID, so an agent can pick up exactly where it left off instead of starting cold every time.
• A helper agent for delegation — open-ended subtasks get handed off with automated checklists, which is the same pattern agencies already use manually when splitting client work across team members.
• Context management that actually controls cost, truncating oversized tool results and triggering compaction once a session hits 85% of its context window, with automatic recovery if it overflows anyway.

That last point is where the framework backs up its claims with numbers. In AWS's own benchmarking on Terminal-Bench 2.1, using the same underlying model (Anthropic's Fable 5), Strands Harness completed tasks for $56.29 on average at 69.7% accuracy, versus $248.05 at 61.8% accuracy for Claude Code running the identical model — a roughly 77% cost reduction with better accuracy, not a trade-off between the two. Independent coverage from MarkTechPost and SiliconANGLE both ran through the same benchmark data, and AWS also claims a 26% efficiency edge over comparable frameworks like OpenCode and oh-my-pi using identical models.

How to Actually Use It

Getting started is a few lines of code:

from strands_harness import create_harness

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

For teams that want to skip writing code entirely, the Strands CLI lets you prototype an agent's behavior in plain English first, then exports working, production-ready code as a Python or TypeScript project you can drop straight into a deployment pipeline. Support for external tools also comes standard through the Model Context Protocol (MCP), so anything you've already built or connected as an MCP server — internal APIs, CRMs, scraped data sources — plugs in without extra glue code.

Who Should Care

If your agency is building agents that need to survive beyond a demo call — client-facing research assistants, automation pipelines that run unattended, or multi-step workflows that used to require a human checking in at every stage — Strands Harness is directly relevant infrastructure, not just another framework to bookmark. It's free, it's model-agnostic, and the benchmark numbers suggest it solves the two problems that actually kill agent projects in production: runaway cost and context loss over long sessions.

Key Takeaways

• AWS released Strands Harness, a free, open-source (Apache 2.0) agent framework that runs on any cloud or locally — not locked to AWS.
• It supports Anthropic, OpenAI, Google, Bedrock, and local Ollama models, so switching model providers doesn't mean rewriting your agent.
• Benchmarks show roughly 77% lower cost than Claude Code on identical tasks and models, with higher accuracy, largely from smarter context management.
• Persistent, resumable sessions and a built-in delegation pattern address the exact reliability gaps that stop agent prototypes from becoming client-ready products.
• Get started with pip install strands-agents-harness or explore the code at github.com/strands-agents/harness-sdk.