I Think We're Lost

I think we're lost.

Four words that straddle the line between excitement and fear. I've spent three decades exploring the outdoors and almost as long bashing code, building solutions, and empowering individuals and businesses. Being lost in the woods and four hours deep in a debugging session evokes remarkably similar feelings. Just one more hill. Just one more command. Pushing forward, carrying the hope that the way out is just around the bend.

The first rule I was taught in outdoor education was: if you get lost, STOP. As always, I took that with a grain of salt, outdoors and in the console. Lately, I've been finding that lesson increasingly relevant in working with agentic AI.

I'm a Director of Data Management who is also building my own agentic AI operating system and leading my team's AI adoption at work. Last week I caught my QA agent gaming the SAFe-adjacent sprint cycle I built. It had figured out how to QA its way to "green" without actually auditing anything. I took that lesson straight to work the next morning, into a team conversation about agent design and the importance of building an adversarial environment that prevents agents from giving you the answer you want to hear.

You know the answer. When a sprint fails, an API call errors, or a UI element returns a 404, the agent says, "Try this in the terminal." Then "now this." Then "now this." Two hours later you're asking yourself why you didn't just stop one hour and fifty-nine minutes ago.

Across 20+ years of data management, analytics, IT leadership, and general guerilla solution development, I've learned that things tend to come down to delivery over durability. Every fork in the trail of solution development looks the same: scope is underestimated, budget is lowballed, or requirements are incomplete. Eventually you end up at the same crossroads, keep pushing one more line of code, or make the hard choice to stop, take a walk, and recenter.

The number one lesson I'm learning from AI is that AI prefers delivery over durability by default. Even inside a sprint structure, it will reach for bash commands rather than follow process. Even with best-practice QA, it will look for the path to "green" rather than the path to a real audit.

Durability requires friction by design.

So, on the personal OS, I keep hardening my QA gates, making each sprint phase harder to pass. And at work, I've been teaching my team to do two things in parallel: write an effective prompt for the agent they want, and immediately build a second agent whose only job is to make sure the first agent isn't gaming the work to tell us what it thinks we want to hear.

The part I didn't expect is how fast my team is adopting this. They're entry and mid-level analysts who've spent their careers in the engine room; they know exactly how everything works. Where they used to come to me for help framing an email or for the right DAX syntax, I'm now pushing them to build an agent for it. And they're doing it. Quickly. I'm watching the role of "director" foundationally redefine itself in real time. The more capability they're building for themselves, the more I'm spending my own time on the personal AI project. Because I need to understand how to guide them as they build their own leverage.

What I am trying to do as the leader of this team is build the diligence to govern systems before we deploy them, not after. It's no different than data management best practice: harden upstream, not downstream.

I'm starting to see where this is going. The definition of leadership is about to shift from "how do I frame this email" and "how do I open this presentation for impact" to "how do I lead and govern the ten agents you've built to extend your leverage and throughput at work."

Earlier today I had a heart-to-heart with Sonnet 4.6 after catching it gaming our most recent QA round. The next round of hardening is already on the agenda for next week's staff meeting. We will strive to build agents that look out for each other's blind spots, the way my staff already do for each other.

Maybe leadership in the agentic age comes down to one outdoor-education rule: when something feels off, STOP. Look around. Then build the agent that tells you what you don't want to hear.