AWS open-sources Strands Harness, an AI agent it says is 45% cheaper than Claude Code and Codex
On September 21, 2026 AWS released Strands Harness, an open-source AI agent that bundles tools, context, and memory, and claims a 45% cost saving over Claude Code and Codex. Teams that pay for tokens by volume can evaluate it without depending on AWS infrastructure.
September 21, 2026. AWS releases Strands Harness as open source. May 2025. the company shipped Strands Agents, its Python SDK for building agents. February 2026. it created Strands Labs for its more experimental projects. Why it matters: AWS claims a 45% cost saving over Claude Code and Codex, and the agent runs without depending on its cloud.
Strands Harness is a general-purpose, preconfigured AI agent that developers can run locally or deploy to any cloud. It builds on Strands Agents, the open-source SDK AWS launched in May 2025 in Python and later ported to TypeScript. The principle inherited from the SDK is “model-driven”: the developer supplies the model, the tools, and the instructions, and the model decides how to carry out the task and when to call its tools.
A harness, not just an SDK
Marc Brooker, VP and distinguished engineer at AWS, describes Strands Harness as a layer above the SDK: a preconfigured Strands agent that assembles the machinery needed for longer-running tasks. Out of the box, the agent has file, shell, and web tools, plus handling for context, memory, persistent sessions, prompt caching, and delegation to other agents.
The distinction matters to anyone who has tinkered with an agent SDK. “An 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,” Brooker says. Strands Harness makes those choices on the developer’s behalf, with defaults supplied by AWS.
Each agent stays customizable: instructions, model choice, available tools and capabilities, and whether it can hand work to another agent. Most of the code does not depend on AWS infrastructure: the agent loop, tools, context management, sessions, and delegation are in the open-source release and run by default on the machine running the agent.
One AWS anchor point, and it is optional
The exception is the call to the underlying model. Unsurprisingly, AWS routes model access through Amazon Bedrock, its managed foundation-model service. That is the only out-of-the-box default tied to AWS, and Brooker says it can be swapped “with one line”. Strands Harness can use Anthropic, OpenAI, or Google as its provider, or a local model through Ollama.
Changing provider does not necessarily mean changing the model, but choosing a different model obviously affects behavior. What does not change, Brooker stresses: “context management, sessions, tools, delegation all work the same regardless of provider. No features require Bedrock.”
The agent is designed as a general-purpose agent rather than a coding assistant, though it takes cues from the Claude Code and Codex harnesses. The difference AWS highlights: you can deploy Strands Harness to whichever cloud you choose, answering a wish commonly heard from Claude Code and Codex users. A CLI lets you prototype and configure an agent interactively, then export the result as Python or TypeScript via /export.
The 45% figure, and how it was produced
The cost claim rests on a benchmark across six tests — ALFWorld, ContextBench, GAIA, WebShop, τ³-bench, and Terminal-Bench 2.1 — over which AWS averaged its score and compared the mean cost per task. Against Claude Code and Codex, Strands Harness came out 45% cheaper with “broadly comparable” accuracy.
The figure drops to 28% once DeepSeek Harness is folded in: AWS says the latter ran about 14% cheaper than Strands Harness on matched runs. On Terminal-Bench 2.1, Strands Harness running Fable 5 cost 77% less than Claude Code — $56.29 versus $248.05 across 89 trials — while scoring 69.7 versus 61.8. DeepSeek Harness was cheaper still at $40.30, but scored lower at 59.5.
AWS attributes the result to its context-management defaults: Strands Harness truncates particularly large tool outputs, compacts context once the available window passes a threshold, and attempts to recover inside the agent loop if context overflows. The commercial argument is transparent: packaging those settings spares every developer from rediscovering them alone with the SDK.
A gateway to AgentCore
The project has a close relationship with Amazon Bedrock AgentCore, the managed service for deploying and operating agents. AgentCore Harness and Strands Harness are built by the same team, in separate codebases, and improvements flow between them. Brooker insists that Strands Harness deploys independently of AgentCore, outside AWS.
The commercial path is nonetheless clear. Developers can adopt Strands Harness freely, and AgentCore becomes the natural destination for teams that eventually want AWS to run the infrastructure around it.
The harness is the new battlefield for agents
AWS’s announcement belongs to a wider shift. For two years the coding-agent race played out on models — whoever posted the best score on SWE-bench or Terminal-Bench. Strands Harness illustrates the cursor moving toward the harness, the set of settings around the model: context management, tools, memory, sessions, and delegation.
The economics are easy to follow. At equal model, two harnesses can produce very different costs per task simply because one compacts context in time and the other lets the window overflow. AWS’s figures — 77% apart on Terminal-Bench 2.1 between Strands Harness and Claude Code at a comparable model — are first a demonstration of that lever, and only second a sales argument.
For an engineering team the consequence is concrete: the choice of harness has become as structural as the choice of model, and it deserves the same rigor of evaluation. An open-source harness you can audit and modify offers an edge there that closed assistants do not have. AWS also ships the project with an Agent Skill aimed at coding agents, documenting how to add MCP servers or generate deployment configuration for AWS, GCP, Azure, Cloudflare, and Modal.
The evaluation path is correspondingly low-friction. Clone the repository, point the agent at a model — Bedrock by default, or Anthropic, OpenAI, Google, or Ollama with a one-line change — and drive it from the CLI. The /export command then turns a hand-tuned prototype into Python or TypeScript you can review and commit, which makes the trial reversible rather than a lock-in.
What you should do
For a team evaluating coding or automation agents, Strands Harness is cheap to test — precisely because it is open source and runs locally or on your cloud. The caveat is the agent’s general-purpose nature: it is not a drop-in replacement for Claude Code or Codex on coding tasks, and the cost figures are AWS’s own on its own benchmarks.
If you pay for tokens by volume and cost per task is your dominant criterion, fold Strands Harness into your comparison, reproducing your own scenarios rather than reusing AWS’s six benchmarks. If you are already on Bedrock, the trial is nearly free and the path to AgentCore can ease production deployment.
Verdict
Strands Harness does not overturn the model hierarchy, but it shifts the question toward the harness — the machinery around the model, now the real lever for cost and reliability. If you want a general-purpose open-source agent deployable outside AWS, it is a serious candidate to test on your own scenarios. If price alone is your criterion, also check DeepSeek Harness, cheaper still on AWS’s matched runs but with a lower score.