A coding agent that edits files and runs commands needs an oracle. Claude Code can inspect a repository, edit files, and run commands and tests [9]. What it cannot supply on its own is what correct looks like for a memory panel, and that is what the specification was for. It had to fix, before any code existed, that application, wired, compressed, cached, committed, available and swap are seven separate numbers rather than one usage figure [6], and that the process view sits next to logical-processor and NUMA-topology views [5]. Those are the decisions that separate a system monitor from something that merely looks like one, and they take as long to work out no matter how fast the prompt runs.
The architecture follows from that. A shared C++ core aligns metric definitions, actions, models and serialization, and three front ends render it in Swift and AppKit, Win32 controls, and Qt 6 [4]. Plummer's own framing is that this costs more platform-specific work than wrapping a web interface would [15]. It also makes generated code checkable: if committed memory is defined in one place, a wrong number on Linux is a front-end bug rather than three codebases disagreeing about the definition.
Then the schedule. Plummer said the first working versions were generated in roughly four and a half hours, followed by testing and refinement [8]. The macOS beta went out on July 31 and the Windows and Linux builds on August 23 [2], which is 23 days [1], about 552 hours of calendar against 4.5 hours of generation, near 123 to 1 [2]. That calendar covers two additional platforms and the hospital stretches Tom's Hardware described [7]. The generation itself was the easy part; the specification and the review around it took the rest of the time.
Three conditions have to hold for those 4.5 hours to transfer to anyone else's project: you have to be able to state correct behaviour before the code exists, failures have to be detectable by running the thing, and you have to recognise wrong output on sight. That third condition is exactly what the account credits Plummer with supplying while Claude Code supplied speed [10]. Remove it and the agent gets you to a wrong answer faster.
Two numbers here are worth checking rather than accepting outright. The 60 frames per second meters are advertised, not independently measured [11], and a monitor drawing per-core graphs at that rate competes for the CPU it is reporting on, which is the constraint the design exists to survive [12]. Separately, the downloads are unsigned, so users on all three platforms may have to approve the application manually before first launch [13]. That approval step is the one part of the process still waiting on a human, agent or no agent.
The lineage matters to the argument. Plummer wrote the first Task Manager at home, got it into Windows NT 4.0 in 1996 after Dave Cutler approved its inclusion [1], and a later Microsoft engineering post still called one of its interface components "Dave's frame class" [14]. The 107 pages are where that judgment went before the agent started typing [7].
