翻訳待ち：Control Charts Make AI Agents Cheaper, Less Necessary, and More Useful

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

According to my new control charts, it’s been 35 days since I posted to this blog. Subjectively, that’s too long. Objectively, it raised an alarm in an email I get every morning. I have long wanted a better and more holistic picture of “how things are going” but instead of settling for an agent like many folks seem to have done, I applied Claude Code to a goal I have had for much longer than agents have existed: Building control charts that show aspects of every part of my life. And while I did not set out to build an agent, I have all these great sources of data in an agentic surface now. The control charts give both me and Claude a frame of reference to discuss, and the context makes Claude a whole lot more useful to me. Honestly, the infrastructure to connect all my information and coalesce it into a useful series of charts is the hard part. I have wanted to do this for years but haven’t had the time to write the code myself until I got my hands on Claude Code. Claude is exceptionally good at the sort of API-glue and Python charting required to bring my vision to life. It found services that I would have had to hunt for — a Plaid-like banking connector that costs very little, Garmin data workarounds, Apple Health export formats. I don’t know how long this would have taken me to build because I know that I wouldn’t have finished building it, but now that I have it, I realize that my highly-measured life had until now been trapped in an archipelago of dashboards. What’s a control chart? Control charts let you visually monitor systems under management. (The system under management is me.) They measure observations over time and raise an alarm based on heuristics. I learned about them in business school and they appealed to my software developer brain — operations management was generally very satisfying, but most of the case examples for control charts come from manufacturing. In its most basic primer, however, the W. Edwards Deming institute says the numbers can be "from anything." Challenge accepted. My control charts use the mean of moving ranges to establish upper and lower natural process limits as described by Donald Wheeler. Conventional implementations use standard deviation (sigma) to draw natural process limits around a mean, but Wheeler argued that this approach bakes the signal into the noise. Claude liked Wheeler’s approach better and I liked Claude’s reasoning. The point of a control chart isn’t to make you hyperaware of how things are going, it’s to let you develop confidence quickly so you can move on, and investigate only the parts of the system that actually cause a problem. For example, my Garmin watch tracks sleep quality, and first thing in the morning, it likes to tell me how I’ve slept. If the answer is “not great,” it can be a distracting and demoralizing way to start my day. Some would say that self-rating is more important, and I generally agree. Garmin rates sleep on a 100-point scale that then gets applied to a “body battery." Your battery depletes based on how you use your body during the day. I have felt fine when my body battery says my sleep was worth 40 points, implying a restless night, and crappy when it starts the day at 90, implying great sleep. As a runner, how I feel while I’m running is much more important to me than what the heart rate on my watch says, in the moment. But the data I collect on the run is very useful to long-term goal-setting and planning. Control charts give me permission to ignore most short term heart rate spikes and watch beeps as insignificant, knowing that the charts will alert me to a more important issue, if one exists. What does this have to do with agents? These control charts solve a whole class problems that I had envisioned an agent solving for me, and they use no tokens unless I want to discuss them further with Claude. If future token economics change dramatically, a million other personal agents will go dark, but my control charts will still be humming along because the job that updates them each morning requires no LLM in the loop. What this really has to do with agents though is all in the context. While I don’t invoke an agent every day, the Claude Code workspace that started with building the control charts has morphed into an ongoing discussion about how things are going. When I’m thinking about whether to reschedule or restructure a training run, Claude can project my fitness and fatigue scores into the future, consider my goals, and give me a helpful response. I could move it to another chat, build a skill, build a chat layer into the dashboard app, but why? This is just for me, and Claude Code is right there in the Claude app on my phone or desktop. There’s a whole separate set of questions implied here about the opportunities and challenges of using LLMs for endurance training, but I suspect there are aspects of what I’m doing that mesh with some of the more thoughtful AI training approaches out there like Mind of Matt from Matt Fitzgerald which lists "intensity balance" right on the top of its list of features. I’ve read a lot about how agents have misled runners into overtraining and injury and am mindful of the risks here. Thus far, the control charts seem to be defining constraints that keep Claude and me both in line. One control I have for running is what percentage of my running time each week is in Zone 1 or Zone 2. This keeps me from overdoing harder sessions without a commensurate ramp in volume — but not too much of a ramp, because that would cause an alarm in my fatigue chart. Maybe you need an AI agent, maybe you don’t — but either way, you might like a control chart.