AWS claims its pre-assembled agent harness cuts costs 28% on the same Claude and GPT models

The loop is already written. On most teams shipping an agent this quarter it is a while-loop around a model call, a tool dispatcher, a truncation rule for when context gets long, and a file on disk standing in for memory.

The Strands harness replaces those parts. It provides the shell, file and web tools in a form the model can reason across, and it automatically loads any skills present ^[4]. It keeps long-term memory across runs, resumes an earlier conversation when handed a session ID, and passes open-ended subtasks to a built-in helper agent with a checklist for multi-step work ^[5]. Mitch Ashley, practice lead for software lifecycle engineering at the Futurum Group, said AWS is in effect making it possible to create a working agent using one line of code ^[10].

The 28% arrives in two versions. The summary version is 28% lower costs across six benchmarks using the same Claude or GPT models ^[8]. Marc Brooker, a vice president and distinguished engineer at AWS, put it narrower: the context window capability offloads bulky tool results to files and to caches that can be reused, reducing the total tokens consumed by as much as 28% ^[7]. "As much as" is a ceiling, so a saving reproduced on other traffic would be 28% or less ^[14]. The devops.com report does not name the six benchmarks or say what the harness was measured against ^[13].

Holding the model constant is the part of the claim a team can use. Nothing about Claude or GPT changed inside AWS's own comparison ^[8], so the same experiment runs on one model: your loop in the first column, the harness in the second, one shared task set across both. Count input tokens per completed task, not per call. A harness that makes three cheap calls where yours made one expensive one improves the per-call figure and leaves the monthly invoice where it was.

Whether the saving shows up at all depends on where your tokens sit. If the same tool output gets re-read into context on every turn, offloading it to a file and caching the reusable part is aimed precisely at that spend ^[7]. If the cost sits in one long generation from a short prompt, context management has little to take out.

So the grid is two questions. First, is your token bill dominated by repeated tool output or by generation. Second, are you maintaining one agent or dozens, because devops.com argues organizations may soon have thousands of agents to build, deploy and continuously update, and that most existing software engineering workflows were not designed for that scale ^[16]. Repeated tool output plus many agents is the box where adopting someone else's loop pays twice. Generation-heavy spend plus a single agent is the box where the code you already wrote is cheap to keep. Brooker said the goal is to make it simpler for developers to build an agent that automates a task unique to their organization, and that DevOps teams are now building and deploying these agents much like any other workload ^[12].