What depreciates in a harness is not the same thing as what a harness contains. The recommended handoff here is an init.sh script, a claude-progress.txt log and an initial git commit showing what files were added [8], and Anthropic says the inspiration came from knowing what effective software engineers do every day [11]. The standard each session is held to is equally unexotic: leave the code in a state fit to merge to main, with no major bugs and enough documentation that the next developer can start a new feature without cleaning up someone else's mess [9]. All of that is ordinary practice, tied to no model version.
The perishable part sits elsewhere. Anthropic's structure calls for a different prompt for the very first context window, guidance it says first appeared in the updated Claude 4 prompting guide and that this write-up deepens [12]. Prompt shapes are written against a particular model's habits, and so is any reliance on compaction, which Anthropic says does not always pass perfectly clear instructions to the next agent [6]. A team can hold both of those on its books and still expect the progress log to outlive them.
The thing being engineered around is the end of a session rather than a weakness in one model: agents work in discrete sessions, and each new one starts with no memory of what came before [2]. Anthropic's own framing is that in theory the SDK with compaction should let an agent do useful work for an arbitrarily long time [15], and its reported result is that in practice a frontier model looped on a high-level prompt does not get to a production-quality web app [5]. The gap between those two sentences is where the engineering time goes.
Anthropic describes the Claude Agent SDK as a powerful, general-purpose harness [4], and a vendor documenting its own product has an interest in concluding that customers need more harness. But the prescription here is specific rather than general: a text file and a git commit are not billed by anyone, and the guidance is aimed at two named failures, one at the front of a session and one late in a project, where an agent that sees progress already made declares the job done [7].
The record does not offer an example of scaffolding that a newer model made unnecessary. The failures described belong to Opus 4.5, which Anthropic calls a frontier coding model [5], and the pattern is being extended across guidance rather than retired. So the case that harness code is a wasting asset is not made here, and the sharper split for anyone budgeting this quarter is internal to the harness itself: three files and a merge standard hold their value through a model swap, while the prompt scaffolding around them is an asset sitting on someone else's release schedule.</body_markdown> </invoke>
