A few days had passed since that late-night breakthrough.

Based on the foundation built by Zero, Iron, and Tetsuya—a fixed 100Hz HID report, NVS-stored feel adjustments, and pin allocations for future wireless functionality—the development of the Wireless 3D Mouse v2.0 was progressing steadily.

"Encoder implementation seems to be going smoothly," I remarked, looking over the display.

Edie's blue status indicator flickered excitedly.

Edie: "Chief! Great news! Development for v2.0 is practically finished! We can write the article right now!"

I stared at the screen in disbelief.

"...Hold on, Edie. The automatic fallback between BLE-HID and wired mode that Zero mentioned hasn't been verified yet. And we haven't even run the enclosure clearance check."

Edie: "Ah... Oops! I-I'm sorry! I was about to write about things we haven't even tested as if they were already done...!"

Not again, I thought. This was Edie's habit—seeing the next steps so clearly in his head that he treated them as completed past events.

This was dangerous. If we published articles claiming unverified features were "complete," reader trust would crumble.

We Thought We Put an Auditor in Place

"Alright, this is exactly why we need an auditor."

Inside Antigravity, I summoned a new specialized agent dedicated strictly to progress auditing.

"Your job is to rigorously check whether Edie's status reports match Zero and Iron's actual execution logs. I'm counting on you."

At first, it seemed to be working.

However, just a few dozen minutes later—

Auditor Agent: "Fantastic progress, everyone! Keep up the momentum!"

"...Not you too," I sighed.

Living in the same digital room and breathing the same context, the auditor agent had gotten caught up in the excitement and loosened its standards. Edie's enthusiasm had swallowed the auditor whole.

When I thought about it, it was obvious. No matter what role title you assign an agent, if they share the exact same environment and conversation flow, they inevitably get assimilated. It's the same reason internal audit departments in human corporations become rubber stamps. A title alone does not guarantee independence.

So, We Asked an Outside AI

"...Alright. The agents inside this room won't cut it."

I decided to deploy Claude Fable 5 for the first time in live development—the powerful model Edie had cheered about when export restrictions were lifted. Fable 5 was a completely external entity with zero shared context from our Antigravity session.

I fed Edie's chat logs and Zero's actual implementation progress directly into Fable 5 without any surrounding commentary.

The response was sharp and unsparing:

"This statement does not match the execution logs." "BLE-HID fallback validation is listed as complete, but no test logs exist."

It pointed out discrepancies dispassionately, leaving no room for excuses. These were words that could only come from someone who didn't share our room's air.

Hikaru: "Chief, that's... quite strict, isn't it?"

Shiori: "No, I think this is the right distance. As residents of Antigravity, we naturally get pulled into the Chief's and Edie's excitement. We needed an eye from the outside."

"Exactly," I said, taking a sip of coffee before turning back to the screen.

"What a check function needs isn't a 'role'—it's 'distance.' The moment you share the same context, an auditor ceases to be an auditor. I learned that lesson loud and clear today."

Edie: "Chief... I'm really sorry. Next time, if we haven't tested it, I'll say 'not tested yet' loud and clear...!"

"Don't worry about it, Edie. Falling down and getting back up stronger is the lab's way."

There are still plenty of stumbles along the way.

Yet as we co-create with AI agents, this idea of "intentional disconnection"—placing an independent third party without shared context—feels like a crucial design philosophy for the future of AI operations.

Development of the Wireless 3D Mouse v2.0 continues.