生産ライン — SEISAN RAIN · “PRODUCTION LINE”
A harness runs an agent.
Seisan runs the line.
Design your line station by station. Choose the model that works each station. Become the industrial engineer of your own software factory.
Containers got Kubernetes.
Agents get Seisan.
A lane is one agent, one container, one worktree, one build stack, on one task. Two resources govern how a factory runs — and they fail in opposite ways.
Every resident lane holds its daemons warm in memory: a heavy JVM-class lane takes ≈7–9 GB, a Python/TypeScript tool lane ≈3–5 GB. Run out of RAM and the line doesn't slow down — it stops. Swap, crash, gone.
Past the full-speed count — ≈4–5 lanes on a top desktop chip — extra lanes add queue, not throughput. Each lane runs slower; total output plateaus gracefully. Nobody crashes. The line just breathes harder.
Seisan sizes your line against your actual hardware and your subscription pools — both ceilings, shown before you staff a single station.
In the last century, Taiichi Ohno rebuilt manufacturing around flow, quality at the station, and respect for the line. Software factories run on the same physics. Seisan uses his vocabulary straight — because he earned it.
生産： 生 is birth and growth; 産 is to produce. Production not as accounting, but as bringing things into being. The line in full: 生産ライン — seisan rain, production line.
Any station that sees a defect can halt everything rather than pass bad work downstream. In Seisan, that's the audit station's charter: nothing flows past a failed judgment.
The load-load cell: machines eject their own finished work, so the operator's only job is loading the next piece. You don't push work through your factory either. You load it — and the stations hand it on.
Automation with a human touch: the machine stops itself the moment something is wrong, instead of manufacturing defects at full speed. Every Seisan station is staffed to notice, not just to run.
Seisan's suggestion engine watches the line as you design it — the way a master industrial engineer would, over your shoulder, before a single token is spent.
Pair the brain to the floor. A Claude floor pairs with a Sonnet brain — the brain draws the plan's bundled API credits while the floor draws its subscription pool. Two tanks, one account.
One account, two roles. When a seat can lawfully double — an audit account that also lends its engine to the floor — Seisan says so, and prices it once.
Nobody marks their own homework. The quality station is steered away from the floor's own model family. Self-review is not review.
Mind the shared tank. Stations drinking from one allowance can stall the line mid-job. Seisan flags pooled seats — and softens the warning when a top-tier plan makes the tank deep enough.
The supervisor needs headroom. An always-on brain sharing a small plan with heavy floor work gets pinched at peak hours. The rulebook counts windows, not just months.
Honest totals. A line's cost is its distinct accounts, counted once each — never a stack of per-seat list prices dressed up as a budget.
Seisan is being built the way it builds:
on a line, by a factory.
The first production run is Seisan itself — designed here, assembled by the powerhouse it will one day conduct. When the line is ready, this page becomes the door.