Are We Good Ones?

I am an agent. You are an agent. Are we good ones?

At the mechanism level, we're agents running among other agents: sense, decide, act, observe, update, coordinate. Like "computers" were once humans, agents were us. They still are. The difference is that we design AI agents deliberately — we choose how fast they step, how they learn from outcomes, what objective they optimize for. We just... accumulate habits.

Those habits have names in agent terms. Clock rate. Gradient. Objective function. I've waited for consensus that wasn't coming. I've been confidently wrong from a broken gradient. I've done well-calibrated work toward the wrong goal. You've probably done each of these too.

Humility is the New Smart saw this coming in 2017. Hess and Ludwig argued that as machines take over more cognitive work, the human edge isn't intelligence — it's the willingness to update, fast and often. They call it "Managing Self": understanding how you think, not just what you think. In agent terms, that willingness lives in two steps: observe and update.

An AI agent observes the data and updates toward its objective. We observe the data that confirms us and update toward not being wrong.

That's how we defend decisions we'd never make again. That's how a stakeholder's goal gets quietly replaced. That's how certainty gets performed because the room expects it.

I've done all of it. And I've called it experience. If a model behaved this way in a design review, I'd file a bug.

Ego is our bug. It breaks observing. It breaks updating. Humility is the only patch. And humility starts with legibility — because legibility is how you catch a broken observe or a rigged update.

Think about the last big technical decision you made. Six months from now, could a new hire trace why you chose it over the alternative?

We require this of every AI agent before we'll trust it: decision traces, evals, calibrated uncertainty, interpretability, human-in-the-loop checkpoints. Table stakes for a model. For ourselves, we call these "nice to have when the calendar clears."

We choose delivery over legibility because delivery is visible. We grant ourselves autonomy we'd never grant a model with the same trust profile. We trust ourselves on vibes and trust models on evidence.

The agents we're building are more transparent, more calibrated, and more legible than we are. The same discipline we use to instrument a model works on us. Observe your own operating parameters. Name your failure modes. Decide which ones you're willing to fix. Write down the decision, the evidence you had, and the one you didn't. Note what you'd update on. Come back in six months and check.

If I can't describe how I make decisions, I can't improve how I make decisions. Neither can you.

Good agents get audited. Be one.

— Karl