AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex

Amazon Web Services has open-sourced Strands Harness, a general-purpose AI agent offering developers a ready-made foundation to run locally or deploy in the cloud. It went live on Monday.

What is inside the Harness

Strands Harness builds on Strands Agents, the open source Python SDK AWS introduced in May 2025 and later brought to TypeScript. February 2026 added Strands Labs, a home for experimental work. Marc Brooker, VP and distinguished engineer at AWS, tells The New Stack that the Harness sits above the SDK and hands developers a preconfigured Strands agent.

The SDK "gives you the building blocks, but you still need to decide how to manage context, persist conversations, integrate tools, and guide the agent's behavior," he says. Out of the box the agent ships with file, shell and web tools, plus context handling, memory, persistent sessions, prompt caching and delegation to other agents; developers can add their own tools and skills and connect MCP servers.

Most of the Harness does not depend on AWS: the agent loop, tools, context management, sessions and delegation are in the open source release and run by default on the machine executing the agent. Only the model call goes through Amazon Bedrock, the single AWS-tied default; Brooker says one line swaps it for Anthropic, OpenAI, Google or a local model via Ollama. "No features require Bedrock."

The cost comparison

AWS says the harness around a model changes cost and accuracy materially, even with the same model. Averaging scores across six benchmarks - ALFWorld, ContextBench, GAIA, WebShop, τ³-bench and Terminal-Bench 2.1 - against cost per task, AWS puts Strands Harness 45% cheaper than Claude Code and Codex at broadly comparable accuracy. Including DeepSeek Harness, about 14% cheaper on matched runs, cuts that to 28%.

On Terminal-Bench 2.1 alone, Strands Harness running Fable 5 cost $56.29 versus $248.05 for Claude Code across 89 trials, 77% less, while scoring 69.7 against 61.8. DeepSeek Harness was cheaper still at $40.30 with a lower score of 59.5. AWS credits context defaults that truncate large tool outputs and compact the context past a threshold.

Why it matters

AWS frames Strands Harness as a general-purpose agent rather than a coding assistant, though it takes cues from Claude Code and Codex; developers can deploy it to whichever cloud they choose. Amazon Bedrock AgentCore, the managed agent-hosting service, was built by the same team in a separate codebase; Brooker stresses it is optional hosting and "all deployment paths are open."