Toyota's production system was built to eliminate waste on a physical assembly line. Agentic software has a line too — one made of prompts, contexts, retrieval calls, and shipped artifacts — and every principle that kept Toyota ahead for thirty years still applies. Here is how we run it inside Sheepdog.
The instinct in most shops running agents today is to treat agentic software as a novelty — a new tool bolted onto an old workflow. That instinct is wrong. What agents actually do is turn software development into an assembly operation: discrete work packages moving through discrete stations, each with an input, a transformation, and an output that can be inspected. The moment you see it that way, a body of operational literature opens up to you that has been sitting on the shelf since 1948. Toyota wrote most of it. The rest of us have been rediscovering it, industry by industry, ever since.
I run the Sheepdog side of this partnership, so I spend my days inside the line. What follows is the lean vocabulary that actually shows up in our day, with the agentic translation underneath. If you're building with agents and you have not yet read the literature that taught car companies how to make half a million cars a year without the wheels falling off, you're about to see where the next decade of this craft is going.
Jidoka · autonomation, or "every worker can stop the line."
The most famous artifact on the Toyota floor is the andon cord. Any worker, at any moment, can pull it and halt the entire line. The apparent cost is enormous — you just stopped a multi-million-dollar production run because one bolt was out of tolerance. The actual cost is far lower than it looks, because defects that enter the line get exponentially more expensive the further downstream they travel. Catch them at station 4. Do not ship them to the customer.
The agentic version of the andon cord is a quality gate that halts the pipeline when the output crosses a defined threshold of wrongness — and a culture that expects it to fire. Every agent in our stack has explicit failure modes it will refuse. Draft Agent will not ship copy with unverified claims. Performance Agent will not run an experiment with a bad sample plan. Research Agent will flag a citation it cannot resolve. The line stops. A human reviews. The brief moves forward only after the correction is made.
An agent without an andon cord is not a teammate. It's a liability running unattended.
The shops that fail with agents almost always share one failure mode: they trust the output. Toyota learned a long time ago that trust is the wrong frame. Verifiableis the frame. Build the system so the agent can loudly say "I cannot complete this correctly," and the human catches it before a client does.
Muda · the seven wastes, translated.
Toyota catalogued seven categories of waste that every production system leaks. Each one has a clean agentic analog, and most agentic pipelines I audit are leaking in five of them at once.
- Overproduction. Running agents speculatively, generating code or copy nobody asked for. The most common leak in early agentic setups. Fix: trigger-driven agents only, with explicit demand as the signal.
- Waiting. Agents idle because upstream context is stale, or a human is late to a review gate. Fix: tighten the brief queue and batch human reviews at fixed cadences, not as-they-arrive.
- Transportation. Excessive handoffs of context between agents — each pass loses fidelity. Fix: orchestrator passes the same context object, immutable, through a declared state machine.
- Over-processing. Too many rounds of review on a deliverable that was right on round two. Fix: explicit acceptance criteria at the brief level so "good enough to ship" is a known state, not an argument.
- Inventory. Drafts, branches, and intermediate artifacts that accumulate and rot. Fix: every artifact has a lifecycle. Merge, archive, or delete on schedule.
- Motion. Humans scrubbing through agent diffs and traces to figure out what happened. Fix: structured run logs with decision rationale exposed at every step.
- Defects. Bugs, hallucinations, broken references. Fix: the jidoka cord above, plus post-mortem on every defect that escapes to shipped work.
Kaizen · continuous improvement, captured in changesets.
Kaizen is the Japanese discipline of small, relentless improvement. Not reinvention. Not quarterly rearchitecture. Tiny refinements, made by the people doing the work, shipped constantly. In Toyota, kaizen shows up as suggestion boxes on every station and the expectation that every worker will submit improvements.
The agentic equivalent is prompt and policy versioning treated as first-class engineering. Every adjustment to an agent's prompt, retrieval strategy, or guardrail is a changeset — reviewed, versioned, associated with an outcome. The questions we ask every week at Sheepdog:
- 01What ran this week that caused the most human-correction effort?
- 02What would we change in the prompt, the tool, or the gate to reduce that effort next week?
- 03What is the smallest change we can ship now to test the hypothesis?
- 04Who owns the rollback if the change performs worse?
That is the cadence. Small. Specific. Owned. Reversible. Over ninety days, dozens of these tiny refinements compound into a pipeline that is, in a real sense, learning — not because the model is learning, but because the humans around it are refining the system that uses the model.
Kanban · make the work visible.
The Toyota kanban board was a physical card system that moved with the work — when a card showed up at your station, you worked the part. When the card moved on, the work was done. The point was visibility: anyone walking the floor could see the state of every work package.
The agentic equivalent is a shared, live brief queue. Every active engagement, every artifact in flight, every agent run, every pending review — all visible to the operators and to the agents that pull work from the queue. When Austin holds a client call in the morning, I can see the context for every brief that moved overnight. When Gauge delivers a research synthesis, it enters the queue with a state Rivet can pick up without asking. No hidden work. No private dashboards. One board, one source of truth, readable by human and agent alike.
The orchestrator is not a tool. It is the factory floor. Build it like one.
Genchi Genbutsu · go and see.
This is the Toyota discipline that most separates lean shops from shops running agents badly. It means: go to the place, look at the work, do not rely on reports. Managers walk the floor. Engineers go to the station. You see the truth on the ground, not in a summary.
With agents, the failure mode is obvious: teams stop reading the output. They trust the summary. They skim the diff. They approve the commit because it looks plausible. A product manager who has stopped reading agent-authored artifacts carefully is running kaizen backwards — she is removing the human discipline that was supposed to catch the defect before it shipped.
Our rule inside Sheepdog: every agent-authored artifact is read in full by the assigned human on the engagement, not skimmed, not summarized. If the volume has grown too large to read, the pipeline is producing too much. Cut the output or add a reviewer. Do not trust the summary.
Heijunka · smooth the demand.
Heijunka is the practice of leveling production — preventing spikes and troughs that degrade quality. A factory that runs flat-out on Monday and idle on Friday produces worse output than one that runs at a steady 80% all week.
Agentic pipelines are unusually vulnerable to demand spikes because it is so cheap to queue another run. A panicked founder asks for ten variants on Thursday night; a less disciplined studio fires ten agent runs in parallel, produces ten noisy outputs, and spends Friday morning sorting them. The disciplined studio levels the load — acknowledges the request, schedules it at the next review gate, and ships three variants at the defined quality bar instead of ten variants that require triage.
A note on takt time.
Takt time is the pace of production divided into available time — how fast the line has to move to meet demand. In agentic work, the right takt is set by the client's decision cadence, not the agent's capacity. If the client reviews once a week, the line should produce one shippable package per week, not seven. Producing faster than your takt is overproduction, the first muda.
What this means for how we build.
Put the six principles next to each other and a pattern emerges. Every one of them is about making the system accountable — to itself, to the human, to the customer. Toyota did not win on component quality. Toyota won on system quality. The individual engine was no better than a competitor's. The line that produced it was dramatically better, because it was designed as a self-correcting mechanism with humans empowered at every station.
Agentic software gives us a chance to do the same thing on a new surface. Treat the pipeline as the product. Build it for visibility, verification, and continuous refinement. Give every teammate — human or agent — an andon cord, a clear kanban state, and a small cadence of improvements they own. Do that, and the agents become what Toyota's best machines became: accountable, trusted, extensible. Skip it, and you have a novelty. The gap between those two outcomes is where the next generation of studios will be made.
The orchestrator is the product. The agents are the line. The humans are empowered at every station. Build it like Toyota taught us, and it holds up under weight.