The Bottleneck Moved

Spent the weekend watching Nate B. Jones' YouTube content — former Head of Product for Amazon Prime Video, now an AI-first product strategist. He articulates a few things I've been feeling but couldn't quite name.

The big one: we're moving from a deterministic world to a probabilistic one.

Software used to be deterministic. You write code, distribute copies, they all behave the same. AI doesn't work that way. Give an agent broad instructions and you're essentially pulling a slot machine. The outcome is uncertain every time. This shift changes everything about how we work.

The work becomes steering

Our job is no longer "write the code." It's steering agentic LLMs toward the outcome we want. The new workflow looks like this:

1. Define clearly what "done" means before you start.
2. Use agentic harnesses to steer and control the probabilistic output.
3. Rigorously check the result at the end.

Simple to say. Hard to do well.

Clarity is the new constraint

For years, execution was the bottleneck in tech. Not enough engineers, not enough sprints, not enough time. AI is dissolving that constraint, but it's not removing bottlenecks — it's moving them upstream.

The new bottleneck is clarity.

Clarity for the org on what customers actually want. Clarity for managers designing new systems. Clarity for everyone defining what "done" means when handing work to agents. If you can't define it precisely, the slot machine just keeps spinning.

We are all managers now

Human authority doesn't go away — it concentrates. While agents can grind for weeks, we become the bottleneck. Our value shifts to taste, testing, and defining the work.

Jones frames it with three questions:

• Can you define the work clearly?
• Can you keep the process unblocked?
• Can you make timely calls about what's correct and what's not?

That's management. Whether you have the title or not.

From roles to skills

Another shift that resonated: we're moving from a role economy to a skill economy. AI drastically reduces the knowledge gap between disciplines. Your job title stops defining what you can do — the skills you actually build do.

The best new developer skill isn't writing faster code. It's gaining clarity on the definition of "done." What can actually be completed in two weeks? What exactly does it look like when finished? We're shifting from execution to delegation, while keeping human authority over what ships.

Some practical bits

A few things from his content that stuck with me as immediately actionable:

• Make the evaluation the outcome. Force yourself to evaluate AI output rigorously — don't just glance at it.
• Explicitly tell it not to lie. Sounds obvious. Works surprisingly well.
• Treat failure as data. A bad output isn't a dead end, it's signal.
• Show unfinished work more often. Steering mid-process is cheaper than fixing at the end.
• Optimize human attention. AI reviewing AI work, with humans intervening only where it matters most. The volume of output will be overwhelming otherwise.

The skill gap is real

If you're feeling frustrated right now, Jones notes that it's normal. We're walking into a space with a huge skill gap. The good news is these are learnable skills — defining work clearly, evaluating output rigorously, steering probabilistic systems. The gap is an opportunity, not a wall.

The bottleneck moved. Might as well get good at the new one.