Stopping my AI agent factory, Toyota-style
My AI agent orchestrator's healthcheck was clean but a dozen PRDs auto-closed and didn't ship.
So I fixed it. The fix wasn't a plaster. It was stopping the production line completely and walking each failure back to a root policy, the way Toyota does on a car line. Jidoka, poka yoke, the andon cord, muda, applied to AI workers that can burn tokens on every cycle.
Work in progress!
Full write-up here: https://blog.mariohayashi.com/p/the-factory-must-grow-part-iii-stopping
What breaks most agent systems is not model quality, it’s silent process debt.
The dangerous part is when the loop looks healthy enough to keep running while quietly producing bad output at scale.
Applying Toyota-style failure discipline here is the right move:
stop the line
trace the failure to policy
fix the system, not the symptom
Most agent stacks don’t fail because the model is weak.
They fail because nobody built a reliable failure doctrine around the model.
That layer is where this gets much more valuable.
If this keeps moving toward agent reliability infrastructure instead of “AI agents that do tasks,” the current framing may age out fast.
Davoq.com would fit that direction better.
Really interesting analogy. I’ve been running an autonomous setup locally (OpenHands on an RTX 4080 Super) to build out an open-core CMS, and the 'factory' mindset is exactly where I got stuck.
We moved to a 'Skills' folder architecture—essentially a source-of-truth directory that our agents must 'read' before starting a task to prevent context drift. It’s the only way we’ve kept the 'Vibe Coding' manageable.
When you say you're 'stopping the factory,' are you moving toward more deterministic logic, or just refining the prompts for better consistency?
Making the agent read a comprehensive Skills folder is definitely the way to go — I use the same pattern too.
When I say stopping the factory, yes, I mean stopping on failure, look for deterministic, structural fix. Deterministic structure first, prompts second. Prompt refinement alone is like whack-a-mole as the model can always hallucinate. What worked was wrapping agents in mandatory pre/post hooks — "forced" reads, output validation, so there's less chance to drift.
Curious how you're using OpenHands though. What are the ins-outs, for someone who hasn't used it before? (Me!)
You mean how I installed it or how it runs? because it runs pretty much the same way any other LLM chatbot works 🤷♂️😅
Andon cord thinking applied to solo founder ops is underrated.
The Toyota insight: quality is built into the process, not inspected at the end. Most solopreneurs run the opposite - ship everything fast, fix problems when clients complain. The result is a constant triage loop that kills focus.
What actually helped me was treating my ops workspace the same way: if something doesn't have a home (client status, project stage, decision log), the system flags it before it becomes a fire. The friction of maintaining structure is much lower than the cost of recurring surprises.