Two of the ideas this station stands on come from other people, and they get their credit before we use them. In 2024, the authors of a widely read essay on a year of building with language models wrote that input–output pairs from production are “the ‘real things, real places’ (genchi genbutsu) of LLM applications, and they cannot be substituted” — the nearest sentence to this entire walk that anyone has published, arrived at from the evaluation side rather than the lean side. And the standing machinery around a coding agent has already been mapped: Birgitta Böckeler’s article on harness engineering — a practice several people helped name in early 2026 — laid out its anatomy, and her map is the one this station builds on. What we found no prior art for is the unification that follows.
Here is the claim the whole street has been building toward. The question Station 4 left us with — where do you stand? — assumes the gemba is one place. It is not, anymore. When the work became a collaboration between people and machines on text, the actual place fragmented into three layers, stacked on top of each other, each with its own kind of visibility, each requiring its own kind of walk. Learning to see them separately, and then to stand in all three, is the discipline everything else in this site serves.
LAYER 1The Transcript
The agent session — the prompts, the tool calls, the retries, the dead ends, the moments a person stepped in — is the closest thing this era has to watching an operator at a machine. It is the work, recorded at full resolution as it happened.
The walk here is concrete: read five transcripts, end to end, weekly. Not summaries of transcripts, not metrics about transcripts — the sessions themselves, the way the manager in the circle watched the press line. The same practice is now recommended, in almost the same words, by an AI lab and by evaluation institutes on two continents, each having arrived from the eval direction; when two traditions that do not read each other converge on “sit down and read the actual sessions,” that is worth treating as signal.
And what does the trained eye see? The same seven shapes it learned at Station 1, wearing new clothes.
An agent thrashing on a task it cannot do is motion — effort without progress, visible only if you watch the effort. Context rebuilt from scratch every session because nothing carries over is transportation, the informational kind: knowledge hauled across the same distance again and again. A human intervention can be value-add (a decision only a person could make) or plain rework (correcting what should not have been generated) — the transcript shows which, and the aggregate never does. And the deadliest read is an absence: the pair that stopped asking the agent anything hard. Kent Beck, watching his own sessions, keeps a list of warning signs in the same spirit — functionality he had not asked for, the agent quietly disabling or deleting tests. The trained eye is the trained eye, whatever the floor is made of.
LAYER 2The Harness
Around every agent stands its harness: the standing instructions, the specs, the gates, the evals (the automated checks that grade what the agent produced) — everything that shapes behavior before and after the model acts. Böckeler’s taxonomy: guides, the feedforward controls that steer the agent before it acts, and sensors, the feedback controls that observe after it acts and force correction. Her conclusion is the one to build on: building this harness “is emerging as an ongoing engineering practice, not a one-time configuration.”
The lean eye adds one observation to hers: the transcript is ephemeral, but the harness compounds. A session evaporates when it ends; a harness rule that prevents a class of waste keeps preventing it in every session that follows. On this floor, the harness is where durable value accumulates — it is the machine tool of the token era, and a factory full of such machines is a subject of its own.
The walk for this layer is a cross-examination: does the harness actually shape the work, or is it aspirational documentation? Take one rule from the standing instructions and trace it through five transcripts. Was it ever load-bearing — did it change what the agent did, catch what it produced? A harness rule no transcript has ever needed is either the fence that is quietly holding, or a sentence nobody reads; only the transcripts can tell you which. And Böckeler points at the harness nobody writes down: the human developer’s own — “we know that our name is on the commit.”
LAYER 3The Flow
The third layer is the oldest: intent to running software, the value stream itself. Here the classical method survives almost unchanged — you can map the stream from “someone wanted this” to “it runs in production” exactly the way Womack mapped steel. What changed is where the stream silts up. When generation is nearly free, the bottleneck migrates upstream and downstream of the typing: to specification (deciding what to want), to review (deciding whether to trust), to decision latency (deciding at all). Building software with AI now means managing exactly those three.
The walk: follow one real feature the whole way, and time where it waited. Not where it was worked on — where it waited. The waiting is almost always for a human decision, and the sum of that waiting routinely dwarfs the generation time by an order of magnitude. One cautionary tale from the practitioner literature: a developer ran an experiment in shipping AI-generated code without reading it. Production broke, and the recovery took weeks. The stream does not care how fast the middle of it flows.
Answering Ballé
Michael Ballé’s objection is still standing from Station 4: you cannot walk a gemba remotely — “lean is a hands-on sport,” and the context that matters does not fit through a screen. About video calls to a physical floor, he was right, and nothing in this station contradicts him: a camera pointed at a factory transmits a picture of the work, minus the heat, the vibration, the operator’s hesitation — minus the context.
But a transcript is not a camera pointed at the work. For work that consists of tokens exchanged between a person and a machine, the transcript is the work — complete, primary, at native resolution. Nothing was lost in transmission because nothing was transmitted; you are not watching a feed from the gemba, you are standing in the artifact itself. Ballé’s principle — context matters, so go to where the context is — survives fully intact. The context moved.
The replayable gemba
One property of this new floor has no precedent at all. Ohno needed the circle because the work vanished as it happened: a missed observation at 9:14 was gone by 9:15. The token-era floor is different — yesterday’s gemba can be walked today, at double speed, paused, rewound. The tooling exists everywhere already, built for debugging and compliance and sold as observability. Reframed through everything this street has taught: replay is the new chalk circle — the commitment device that holds you in front of the work until you see it.
It also carries a trap. Replay makes observation feel like a report — skimmable, summarizable, delegable to a dashboard that watches the replays for you. Every convenience of the new circle whispers the old temptation: surely someone could summarize this. Two habits guard against it. Walk the replays on a schedule, the way Ohno’s students stood in chalk. And occasionally sit live beside a developer–agent pair anyway — because the transcript, complete as it is, never shows the prompt that was typed and deleted, or the question that stopped being asked the day the developer stopped trusting the answer.
You now know what the three layers are and how each is walked. What you do not yet know is whether you would actually see anything — whether the trained eye survives the translation. There is a way to find out, and it takes five minutes.