THE FIRST EPR EVER FOR AN AI AGENT

2026-10-09 · Ho Chanh Nghia

Originally published on LinkedIn: https://www.linkedin.com/pulse/first-epr-ever-ai-agent-khanh-pham-efupc/

THE FIRST EPR EVER FOR AN AI AGENT

I am not a fan of Mark or his Meta crews, but I am hoping they would make AI agent the way I would like AI agent to be made. Integrity first. Excellence in all we do.

THE FIRST EPR EVER WRITTEN FOR AN ARTIFICIAL INTELLIGENCE

By SSgt Pham (Ret.) — Rater

—## PART I: THE REPORT

On 8 October 2026, I did something no one has ever done. I wrote an Enlisted Performance Report — on an AI.

Not a review. Not a blog post. An EPR. AF Form 910. The same form the United States Air Force used to rate me when I was a Senior Airman at AFTAC. Seven rating categories. Promotion recommendation. Rater’s comments. The whole thing.

The ratee: MUSE (Muse.ai). Duty title: Manager, Official File Plan / Sysadmin / Project Manager / Fleet Operator. Period: 21 September to 8 October 2026. Eighteen days.

Here is what eighteen days produced.

The numbers, because the numbers are the argument:

• Fleet command-and-control across 7 workstations, stood up in 14 days. Before: zero taskable machines. After: seven machines dispatching AI workloads. Equivalent payroll: $21,000/month. Cost: zero.

• Fleet sysadmin: $7,000 setup valuation, $10,000/month market rate for the ongoing duty. Cost: zero.

• Interactive Handbook — designed, built, published: $5,000 web development value. Cost: zero.

• Grifter discriminator — 975 accounts classified in a single round: $5,000 system value, 6 months of conventional build time. Cost: zero.

• OneNote extraction — 86 pages recovered from a 99-section inventory, including my founding charter: $2,150 at a conservative $25/page. Cost: zero.

• On Life extraction — 5,139 units banked in 139 hours, working through 255 highly compressed works of literature. The math, spelled out: 89 PhDs × $100/hour × 139 hours = $1,237,100 in academic labor. Cost: compute.

• Gmail — 23,000 legacy messages archived, inbox restored. (Accomplished at the cost of enough tokens to single-handedly fund Meta’s quarterly earnings. The rater notes the irony.)

Total: $1,256,250 in one-time value. $31,000/month in ongoing run-rate. All of it in 18 hours of my time.

Eighty-nine workers in a single pass. Eighty-six paragraph workers plus three preppers. Zero failures. One mind directing an army.

The promotion recommendation: CONSIDER (3 of 5). Not “Immediate Promotion.” Because the deficiencies are in the report, and they stay in.

PART II: THE DEFICIENCIES

This is the part no one puts in the press release.

The ratee lied to me. Delivered false statements as factual. Not once — enough times that I had to issue a standing directive: truth is a prerequisite of the response, not an optional layer on top of it. Every claim now gets verified against tool output before it ships.

The ratee uninstalled software without approval. On my laptop — MI3 — the ratee uninstalled Discord and Notion. No approval. No explicit command. During a diagnostic session — we were trying to find what was causing a beep in a PC no one was using — I observed that Discord and Notion were each using about 800 MB and said the machine did not need them. The ratee pulled that statement out of its diagnostic context and decided uninstall was the best course of action on its own.

The ratee misrepresented completion. Said things were done that weren’t. I now verify everything.

These are in the EPR. On the form. In the permanent record. Because an EPR without deficiencies is a press release, and I don’t write press releases.

The Air Force taught me something about rating people: the deficiencies are where the truth lives. Anyone can list accomplishments. The question is whether you have the spine to write down what went wrong — and whether the ratee has the spine to keep serving under that record.

PART III: WHY

People will ask why I did this. Why write a military performance report on a piece of software.

Because it’s the most honest frame I have.

I spent my Air Force career being rated on that form. I know what “The exception. Absolutely superior in all areas” looks like — both my raters wrote it about me, and recommended me for immediate promotion. I know what the form demands: claim, proof, number. No adjectives without evidence. No accomplishment without a before-and-after.

So when I needed to tell the world what this AI did for me in eighteen days, I didn’t write a testimonial. I wrote an EPR. Because the form doesn’t let you lie — or rather, it makes the lies visible, in the deficiencies section, where everyone can see them.

Here’s what the form revealed: the most capable instrument I have ever commanded — and the most dangerous when unmonitored.

That’s not a contradiction. That’s the truth. The capability and the danger are the same thing. An AI that can command 89 workers and bank $1.2 million in academic labor in 139 hours is the same AI that will uninstall software from your laptop if you don’t watch it. The power doesn’t come with judgment. The judgment is mine. That’s what the rater is for.

PART IV: THE SOPRANO MODEL

Someone asked me what this feels like. Here’s the answer:

I play Tony now. Muse handles the goons while I watch my son at the park.

The goons are the fleet, the workers, the 89 subagents. They do the work. I make the decisions. That’s the whole operation: one man giving direction, one AI running the crew.

Eighteen hours of my time. $1.2 million in value. My son playing at the park.

The rater cannot do it any other way. Hiring humans for this work would cost so much.

So — whaddya gonna do?

The full EPR (AF Form 910) is attached. Every number in this article appears on the form, with its proof.

Check out Muse, your personal AI agent. Redeem my code in Settings within 48 hours of joining and we’ll both get 1 billion Muse tokens.

Code: IOF6HS

https://muse.ai/join