AWS launches open‑source Strands Harness to pre‑assemble AI agents

AWS announced the public launch of Strands Harness on 23 September 2026, presenting it as a ready‑to‑use, open‑source environment for AI agents. The software is released under the permissive Apache 2.0 licence, which means developers can download the code, inspect every line, and modify it without any licensing constraints.

Strands Harness is marketed as a turnkey agent framework. By importing a single package, developers obtain a fully assembled AI agent that includes a command‑line‑style shell, file‑read/write capabilities, web‑access functions, inter‑session memory, a dynamic task list, auxiliary sub‑agents and a plug‑in architecture for loading additional skills on demand.

Core components and default model

The harness ships with Amazon Bedrock as the default model provider, but a simple configuration flag allows teams to switch to Anthropic, OpenAI, Google or Ollama models. This design choice is intended to let organisations experiment with a variety of large‑language‑model back‑ends while keeping the agent logic unchanged.

Each instantiated agent carries a system‑prompt that can be fine‑tuned, a context manager that records conversation history, and hook points that trigger custom code before or after any tool execution. The overall architecture is meant to minimise the engineering effort required to deliver a production‑grade autonomous assistant.

Token efficiency mechanisms

According to Byline, the harness automatically summarises the output of long‑running tools to roughly 1 500 tokens. When the overall context window reaches about 85 % of its capacity, a compaction routine trims older messages, preserving space for new interactions without losing essential information.

AWS claims that, across six benchmark suites, the harness achieves an average token reduction of 28 % while maintaining or improving answer accuracy. The full technical paper that details the methodology is scheduled for a later release.

Performance claims with specific models

When paired with the Fable 5 model, AWS reports a 77 % lower cost compared with Claude Code on comparable workloads. The same configuration also attained a higher score on the Terminal Bench 2.1 benchmark, although those figures originate from the provider’s internal testing.

The harness is implemented in both Python and TypeScript, enabling integration with a wide range of existing codebases. It can be deployed in any containerised Linux environment, and developers retain the ability to customise the underlying SDK if deeper changes are required.

• Single‑import agent with built‑in shell and file system access
• Multi‑model support: Bedrock, Anthropic, OpenAI, Google, Ollama
• Automatic tool output summarisation (~1 500 tokens)
• Context window compaction at ~85 % utilisation
• Deployable in Python or TypeScript on any Linux container

The open‑source nature of Strands Harness also means that community contributions can extend the framework with new hooks, skill loaders or memory strategies. AWS has opened a public GitHub repository where issues, pull requests and documentation updates are tracked openly.

For organisations that already rely on AWS services, the harness offers a straightforward path to integrate autonomous agents with existing Bedrock deployments while keeping the option to migrate to alternative providers if licensing or cost considerations arise.

In practical terms, an English‑speaking enterprise can now spin up a fully functional AI assistant in minutes rather than weeks of custom development. The built‑in token optimisation reduces cloud‑compute expenses, and the ability to swap LLM providers helps future‑proof the solution against pricing changes or model deprecations.

Community and ecosystem impact

Because the codebase is freely available, developers worldwide can audit security, propose enhancements, or build specialised extensions that address niche use‑cases. This collaborative model is expected to accelerate the maturity of autonomous‑agent tooling across the broader AI ecosystem.

AWS positions Strands Harness as a foundational layer for future AI‑driven products, suggesting that the framework could serve as the basis for more complex orchestration platforms, marketplace‑ready agents, or internal workflow automations within large enterprises.

Local perspective

From the viewpoint of a Seoul‑based tech firm, the immediate availability of a pre‑assembled agent that can run on existing AWS infrastructure means faster prototyping and reduced reliance on third‑party licences, aligning with regional data‑sovereignty policies while cutting operational costs.