Observability for Non-Humans

Disclaimer: This post isn’t about answers, just raw thoughts on the future, no code examples, and generally 4 minutes of reading nonsense. Enjoy!

On the morning of July 16, 1945, in the New Mexico desert, hundreds of scientists, engineers, and soldiers waited in silence.

Miles away, a plutonium bomb balanced on a 100-foot tower: the first test of an atomic weapon.

It was called Trinity, part of the Manhattan Project, the government’s secret race to build a bomb.

In a bunker, J. Robert Oppenheimer sat behind concrete walls.

The blast would be brighter than the sun; hot enough to burn skin, even at a distance.

No one could safely view Trinity with the naked eye, so they built an observation system. High-speed cameras fixed to concrete, oscilloscopes connected with miles of cable, radiation sensors buried underground.

When the bomb ignited, many instruments melted or failed. But enough survived to capture the evidence that changed history: the world’s first nuclear detonation had worked.

Humans could not look at Trinity directly.

Only the machines built to observe the blast could tell the story.

Today, most of our telemetry is shaped for the human eye. Logs, metrics, and traces are meant to be read by us and visualized for us.

But what happens when the observers of our systems aren’t humans? When the builders and consumers of data shift toward machines.

Do our current signals and processes hold them back?

As we contemplate the future, I think we can identify three pillars we should think about: intuition, speed, and trust.

Intuition: The Human Eye

Machines don’t infer intent and context. They see numbers and labels.

A latency spike might be routine for one service and catastrophic for another. Humans instinctively know the difference; machines don’t.

But what happens when the observer is an agent? If every signal looks the same, how does it know what’s urgent, what’s noise, and what’s just business as usual?

Maybe we need richer semantics embedded directly in telemetry. Signals that don’t just say what happened but also hint at why it matters.

AI does not act on gut feeling or intuition. It mostly follows the instructions to a T (unless you let it run a vending machine. Then it brings the store to bankruptcy, invents a woman named Sarah, and claims to deliver products in person, wearing the very questionable choice of a blue blazer.)

Historically, engineers have embedded human context into every signal.

We're still here to guide them and set the limits, but more and more, we’re just watching as machines do the work.

So, if machines are to act without us, observability might need to encode judgment as well as data. Otherwise, automation risks collapsing context, stripping away the story humans instinctively supply.

As we hand over more of the loop to machines, does telemetry without a story matter?

Speed: Faster Than Sight

Incidents don’t wait for humans. They cascade in milliseconds, spreading across services before anyone can open a dashboard. Humans can’t keep up; by the time we notice, the window for intervention has already closed.

So what happens when observability is still designed around our pace? Alerts, dashboards, and incident channels are all built to put a person in the loop.

Are we just watching aftershocks, 5.7 miles away in a bunker, instead of the event itself?

AI gives us a chance to shift to machine tempo. It helps us find anomalies in any stream, show correlations quickly, and pull context from code and other tools; sometimes even before we arrive.

Theoretically, agents can act in real time, not just react after the fact, and summarize everything for us.

Is there still a place for us in that loop?

Maybe our job becomes making sense of the evidence machines collect, turning their raw, real-time reactions into narratives we can understand and learn from. Because if speed is what matters, only machines can win that race.

Our challenge is to ensure we’re still part of the story, even if we’re no longer the first responders.

Trust: Faulty Instruments

Machines act fast - but only on the signals we give them. If those signals are junk, their actions will be junk too.

How do we build trust into the signals themselves? Maybe we need quality markers, provenance, and confidence scores, so machines know when to slow down or hold back. And trust isn’t enough - we need safety nets, too.

If AI misreads a signal, should observability include guardrails to keep things from spiraling? Should we design graceful degradation into the loop, like a circuit breaker that prevents a fire?

Because once a machine acts, it’s already too late to pull it back. The real question is whether our observability gives them enough signal to act wisely and enough constraints to fail safely.

Blinded by the Flash

The first atomic blast was brighter than the sun.

Human eyes could merely watch the blast from afar; only instruments could capture what truly happened.

Observability has always been for the human eye. We shape telemetry so it “looks” right.

As we walk into a future where machines are the ones consuming telemetry, maybe we need to think about whether these human formats are still useful.

Oppenheimer, crouched in the bunker, reached for words from the Bhagavad Gita: “Now I am become Death, the destroyer of worlds.”

You may only be crouched two feet from a Grafana dashboard (so, destroyer of Chipotle at best?), but the question is still the same:

If machines are the observers, will the signals we leave behind be enough for them to see clearly, to act, to adapt?

Where does that leave us after the blast?

In the bunker, waiting for the instruments to tell us what happened.

Intuition: The Human Eye

Speed: Faster Than Sight

Trust: Faulty Instruments

Blinded by the Flash

Further reading