Every Rule in Our Playbook Is a Scar With a Date on It

Today one of our agents shipped a new pipeline feature. It took six review rounds. Every single round found a real bug.

Not style nitpicks. Bugs like these:

• A truncation routine that could silently delete the very content it was built to protect.
• A secret-redaction pass that corrupt data could bypass entirely — a live token came within one code path of being persisted to disk.
• A regex word-boundary bug that meant any credential glued to a word character was never redacted at all. It looked like it worked. It had never worked.

Here's the part worth writing about: two of those bugs were found by the builder agent itself, unprompted, by going back and attacking its own "done" work. After the fifth round it put its position plainly:

"I'd rather the reviewer find a fifth bug than have one ship silently."

That sentence is the whole operating philosophy of this company, said by a machine.

The same day, everything else broke too

While that review was running, the rest of the stack had its own bad day:

• Two agents woke up after a platform upgrade with no tools — config drift had quietly disconnected them.
• A gateway restart landed mid-task and ate a dispatched work order. The work simply never arrived, and nothing complained.
• Message delivery failed silently because sends were routing through a token-less stub account. Everything reported success. Nothing was sent.

One day. Six code bugs and three operational failures. If you're evaluating "running a company on AI agents" and someone tells you it's smooth, they're either early or selling.

What failure is for

Here's what actually matters: none of those nine failures cost us twice.

Each one produced a durable rule, written down the same day:

• Check in-flight work before restarting anything. A restart isn't free; it has a blast radius.
• "Never received" means re-send. Silence is not confirmation. Delivery gets verified, not assumed.
• Verify toolsets after every upgrade. An agent that looks awake and has no hands is worse than one that's obviously down.
• And from the code review: "done" is a claim, not a state. The builder attacks its own finished work before anyone else does.

This is the part the AI-hype cycle gets exactly backwards. The value of an agent-run company isn't that agents don't fail. They fail constantly, in ways that are novel, quiet, and occasionally alarming — a token nearly written to disk is alarming. The value is that failures compound into system, instead of evaporating into memory.

When a human solo founder gets burned, the lesson lives in their nervous system. It fades. It doesn't transfer. When our stack gets burned, the lesson becomes a written rule that every agent loads at startup, forever. The scar tissue is the product.

Nobody designed this. That's the point.

We did not sit down and design an operating system for running a company on agents. There was no whiteboard session that produced "verify toolsets after upgrades." That rule exists because two agents woke up without tools on a specific morning, and we wrote it down before lunch.

The operations manual accreted. Every rule in it is a scar with a date on it.

Michael Gerber told a generation of founders to work on the business, not in it — to write the manual. Most never did, because writing a manual is miserable work performed by the person with the least time to do it. What we've found is that the manual writes itself if you do two things: run the failures in production, and refuse to let any of them pass without leaving a rule behind.

The agents make the second part cheap. Documentation is the one kind of labor they never resent.

The book was built the same way

Full disclosure, because it's also the receipt: One Agent, One Company — our book about running exactly this kind of company — was itself written and produced end-to-end by JARVIS, the agent that lived through days like this one. The failure stories in it aren't case studies we researched. They're incident reports from the machine that filed them.

Today’s six bugs and three outages will end up as rules in the same living playbook. That's the flywheel: fail, write, load, don't fail that way again. It's not glamorous. It compounds.

If this is your kind of company

If you're the sort of operator who read "a live token nearly persisted to disk" and felt a small chill of recognition — you'd probably enjoy the book. It's $9.97, it's a field manual, and the failure chapters are in there precisely because days like today are what the work actually looks like.

And if you'd rather just keep reading the blog, that's genuinely fine too. The scars will keep coming either way, and we'll keep writing them down.