Amazon's Strands Harness bundles agent tooling into one environment, with reported token savings up to 77% on some benchmarks.
Amazon Web Services released Strands Harness on September 25, 2026 — an open-source framework for building AI agents with less complexity and lower token costs.
The tool consolidates agent development functions into a single environment, letting developers select a model and launch an agent with minimal boilerplate — reportedly one line of code in basic configurations.
Cost savings that matter in production
AWS internal tests across six categories showed 28% average cost reduction using the same models (Claude, ChatGPT, and others) compared with existing agent tooling. Tests with Anthropic's Fable 5 model reported savings up to 77%.
The gains come from how Strands Harness manages context:
- Summarizes lengthy agent outputs instead of re-ingesting full transcripts
- Caches repeated queries to avoid redundant token spend
- Delegates subtasks to sub-agents with scoped context windows
Long conversations are expensive because models re-read entire histories. Strands attacks that structural cost directly.
Features for real development workflows
The default agent can:
- Execute shell commands
- Read, write, and edit files
- Search the web
- Delegate complex tasks to sub-agents with checklist-style progress tracking
Strands CLI, released alongside Harness, supports spoken or written prototyping without writing code initially — then exports finished workflows as code for customization.
Model and deployment flexibility
Strands works with Claude, GPT, Google models, and local Ollama models. Because it runs in Linux containers, it deploys on AWS, Google Cloud, Azure, or on-premises infrastructure — reducing vendor lock-in relative to single-cloud agent platforms.
Benchmark claims
AWS cited improved scores on Terminal Bench 2.1, which measures real-world task completion — suggesting the savings do not come purely from truncating context at the expense of quality.
Independent verification will matter, but the architectural approach — aggressive context management and sub-agent delegation — aligns with known best practices for agent cost control.
Getting started
Developers can explore Strands Harness via AWS's open-source release and Strands CLI documentation. For Stackademic readers building first agent projects, the lesson is foundational: token economics is software architecture when LLMs are your runtime.
Before adding capabilities, design how your agent forgets, summarizes, and delegates. Strands Harness encodes those patterns as defaults — a useful reference even if you build custom frameworks.
Comments
Loading comments…