AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex
The New Stack Paul Sawers ● Covered by 2 sources
AWS open-sourced Strands Harness, a general AI agent devs can run locally or in the cloud. AWS says it can be cheaper than Claude Code and Codex without losing much accuracy.
Based on reporting by The New Stack, Paul Sawers — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon Web Services is opening up a new general-purpose AI agent called Strands Harness, and the pitch is simple: give developers a ready-made base they can run on their own machine or deploy elsewhere. It’s built on Strands Agents, AWS’s open source Python SDK from May 2025, and it sits above that toolkit rather than replacing it.
That matters because AWS is trying to package the messy parts of agent building, not just the model call. Strands Harness ships with file, shell and web tools, plus context handling, memory, persistent sessions, prompt caching and delegation to other agents. Developers can still swap instructions, models, tools, capabilities and handoff rules, but the defaults are already there.
Marc Brooker, AWS vice president and distinguished engineer, says the open source project runs most of that machinery on the same machine by default. The one big exception is the model itself, which routes through Amazon Bedrock out of the box. Even that is not locked in. AWS says developers can switch to Anthropic, OpenAI, Google or a local model through Ollama, and no feature requires Bedrock.
AWS is also making a cost argument. In its own testing across six benchmarks — ALFWorld, ContextBench, GAIA, WebShop, τ³-bench and Terminal-Bench 2.1 — the company says Strands Harness came out 45% cheaper than Claude Code and Codex on average, with broadly comparable accuracy. That figure drops to 28% when DeepSeek Harness is added to the comparison, since AWS says DeepSeek Harness was cheaper still in matched runs.
The biggest reason, according to AWS, is how the harness handles context. It truncates large tool outputs, compacts context when the window crosses a threshold, and tries to recover inside the agent loop if things overflow. On Terminal Bench 2.1, AWS says Strands Harness running Fable 5 cost $56.29 versus $248.05 for Claude Code across 89 trials, while scoring 69.7 versus 61.8. DeepSeek Harness was cheaper again at $40.30, though its score was lower at 59.5.
There’s a commercial angle here, too. AWS also has Bedrock AgentCore, a managed layer for running agents with identity, access controls, observability and hosting. AgentCore Harness and Strands Harness were built by the same team, even though they live in separate codebases. The open source agent is set to go live later on Monday.
My take — AI-written commentary, not fact-checked reporting
AWS is doing the oldest cloud trick in the book: give away the tool, then sell the place to run it. The open source bit is real, but so is the gravity toward Bedrock and AgentCore. That’s not evil; it’s just enterprise software with cleaner packaging and a nicer smile.
Read more about this at: The New Stack