The system runs itself. The organization improves it. My Managing Director comes to me when something needs a decision.

Three ways

I work with AI in three distinct ways at Abyss Applied: augmentation, automation, and agency. They aren't interchangeable, and I don't trust them equally.

I use augmentation to build. I use automation to operate. I grant agency when the evidence earns it.

Some work is deliberately collaborative.

I bring the domain knowledge, engineering judgment, intent, and accountability. AI brings another reasoning surface: challenging assumptions, exploring alternatives, finding gaps, and helping turn ideas into implementations faster than I could alone.

Neither party is simply handing work to the other.

The result is force multiplied through augmentation.

Automation is where I remove myself from work the system has demonstrated it can perform reliably. Research happens. Software gets built. Tests run. Deployments happen. Evidence is retained.

I don't need to be in every loop.

AI can propose. Deterministic code and invariant gates decide. Once proven, the resulting automation runs without either AI reasoning or me in the execution path.

That distinction matters more than the word "automation" usually carries. Handing work to a system is not the same as handing it production authority. A gate either verifies or it fails; nothing proceeds on a probable answer.

Agency is different.

I work primarily through my Managing Director now. I don't directly manage every engineer, and I didn't start by giving the system that authority.

It earned it.

I trained the engineers first. I established their constraints, evaluated their work, corrected failure patterns, and learned where they were reliable. Then I trained the Managing Director to manage them. Then the Managing Director Assistant.

Authority moved upward only as the evidence supported it.

Telemetry is what makes delegation possible. It makes model behavior observable. Once behavior is observable, patterns become measurable and increasingly predictable.

When something fails, I don't automatically change the prompt. I determine where the failure actually belongs — the bot, the code, the infrastructure, or the process. Then I repair it, evaluate it again, and decide whether the system has earned more trust.

That creates the operating loop: observe, evaluate, repair, verify, trust, delegate — and back to observe.

1. 01

   Observe

   Telemetry makes behavior observable.

2. 02

   Evaluate

   Patterns become measurable and predictable.

3. 03

   Repair

   Fix the bot, the code, the infrastructure, or the process.

4. 04

   Verify

   Test again. Confirm the fix.

5. 05

   Trust

   Confidence is earned with evidence.

6. 06

   Delegate

   Expand authority. Increase autonomy.

…and back to Observe.

I built the system. I removed myself from it. Now it operates, improves, and delivers — every day.

So I can work on what moves the mission forward.

Agency isn't something I give an AI because the model is impressive. Agency is earned through demonstrated behavior inside controlled systems.

That is AI fluency as I practice it at Abyss Applied:

• Augment what benefits from both of us.
• Automate what no longer needs me.
• Delegate agency only when it has been earned.