AWS introduced Strands Harness, an open‑source general‑purpose AI agent that adds a ready‑made set of tools, context handling, memory, and session management on top of the existing Strands SDK. The change matters because the agent can be run on a developer’s workstation or any cloud, swaps model back‑ends with a single line, and AWS reports up to 45 % lower cost compared with Claude Code and Codex for the same benchmark suite.
What Strands Harness Provides Out of the Box
The package ships with file, shell, and web interaction tools, built‑in prompt caching, and a delegation mechanism that lets one agent hand work to another. Context and memory are automatically trimmed and compacted when the token window exceeds a configurable threshold, and session state persists across invocations. Developers can still adjust instructions, enable or disable specific tools, and choose the underlying model.
Operational Flexibility and Provider Independence
All core components – the agent loop, tool wrappers, context manager, and session store – execute on the host machine by default, meaning no mandatory AWS services are required. The only default external dependency is the model call, which routes through Amazon Bedrock. That call can be redirected to Anthropic, OpenAI, Google, or a locally hosted model via Ollama with a single configuration change, preserving the same agent behavior regardless of provider.
Cost and Performance Findings
AWS benchmarked Strands Harness across six tasks (ALFWorld, ContextBench, GAIA, WebShop, τ³‑bench, Terminal‑Bench 2.1) and compared average cost per task to Claude Code and Codex. The reported result is a 45 % reduction in cost with comparable accuracy. When DeepSeek Harness is added to the comparison set, the overall savings drop to 28 %. On Terminal‑Bench 2.1 specifically, the harness running the Fable 5 model cost $56.29 versus $248.05 for Claude Code, a 77 % reduction, while scoring 69.7 versus 61.8.
Security and Architectural Considerations
Because the agent’s runtime resides on the host, the primary security boundary is the host environment. Switching model providers changes the trust surface: using Bedrock introduces AWS‑managed data handling, whereas a local Ollama instance keeps data on‑premises. Custom tool integration and session storage locations are also user‑controlled, so teams must assess data exposure for any external tool or storage backend they attach.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams building AI‑driven workflows should evaluate Strands Harness when they need a customizable agent without locking into a single cloud provider. Run the default setup locally to benchmark cost and latency, then experiment with alternative model back‑ends to balance performance, expense, and data‑privacy requirements. Pay particular attention to how context truncation and session persistence are configured, as these directly affect both cost and correctness. Finally, treat the model provider choice as a security decision and verify that any external service meets your organization’s data‑handling policies before production deployment.