Strands Harness
Strands Harness launches as an open-source AI agent solution, promising significant cost reductions and improved performance over existing setups. While the project champions its token-efficiency and ease of use, the Hacker News community dives deep into benchmark skepticism, the crowded harness ecosystem, and AWS's motivations, questioning if this new entry truly disrupts or merely adds to the noise.
The Lowdown
Strands Harness, an open-source AI agent harness from an AWS team, has launched with the ambitious goal of offering 'frontier performance with 28% lower token cost.' Designed for general-purpose agents, it aims to simplify the creation and deployment of AI agents by providing a 'batteries included' solution that works seamlessly across various models and cloud providers.
Key features and claims include:
- An Apache 2.0 licensed, fully assembled agent harness runnable locally or deployable to popular cloud providers.
- Claims 28% lower token cost on average and up to 77% lower with Fable 5, while maintaining or improving accuracy compared to Claude Code, Codex, and other popular harnesses.
- Achieves token-efficiency through sophisticated prompt caching and context management, including tool result truncation, summarization, and context recovery.
- Supports a wide range of models from Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM, with easy Python or TypeScript integration.
- Comes with essential tools like shell, file operations, and web access; manages context windows, long-term memory, and delegates subtasks to helper agents.
- Deployable as a Linux container on platforms like Modal, Cloudflare Containers, Azure Container Apps, and Amazon ECS.
- Includes a Strands CLI for English-based prototyping, allowing users to export generated code.
- Highly customizable, enabling users to override defaults, swap models, add tools, or gradually replace components down to the SDK level.
The team hopes Strands Harness will empower developers to quickly prototype and build more useful agents, inviting community engagement on Discord.
The Gossip
Benchmarking Blues & Harness Comparisons
The benchmarks presented, particularly the use of 'Terminal Bench 2.1,' faced significant skepticism. Commenters questioned its saturation and the specific scores, with some pointing out that Deepseek was cheaper (though less accurate) and asking why popular alternatives like vanilla Pi or Maki.sh were omitted from comparisons. The Strands team responded by clarifying scores and indicating future deep dives into benchmarks and other harnesses.
Third-Party Tensions & Custom Harness Craze
Many users expressed increasing hesitation about adopting yet another third-party agent harness. Concerns ranged from model providers optimizing for their native harnesses (potentially disadvantaging third-party solutions) to the perceived ease of building a custom harness using existing open-source frameworks like Pi or Opencode. Some argued that the initial claim of wiring up primitives being 'tricky' was overstated, as simple delegation patterns can manage costs effectively.
AWS Agendas & Open-Source Optics
The revelation that Strands Harness originates from an AWS team sparked debate about Amazon's broader AI strategy and intentions. Some commenters noted AWS's perceived lag in AI innovations outside of cloud infrastructure and highlighted aggressive sales tactics for AgentCore. While the Strands team emphasized its open-source nature and deployability beyond AWS, some remained skeptical, questioning whether Amazon's long-term goals align with truly open and platform-agnostic development.
Canny Cost Controls & Model Management
Users shared their diverse strategies for optimizing costs and model usage within their agent workflows. Discussions covered leveraging OpenAI subscription pricing, integrating open-weight models via platforms like Ollama, and using services like OpenRouter for specific needs. The conversation also touched upon the complexities of using Anthropic models with non-native harnesses, citing potential TOS violations and recent changes in Anthropic's policy regarding custom harness usage with subscription billing.