A score is a doorway, not an outcome
A source scored five out of five and nothing happened next. The score said where to look, not that anyone had looked.
Insights
Field notes from running governed AI.
Short, practical write-ups drawn from operating a governed AI environment and from modernization work on systems where mistakes are expensive. Each lesson is paired with the failure that taught it.
Where does the human choice become data?
An approval step read the newest version, not the one I chose. The human was near the loop, not in it.
The harness is the work
The prompt tells an agent what to try. The harness decides what it can reach, which checks run, and who can accept the result. Five outside voices, and where my own fleet is taking it next.
The citations were right all along
Three notes cited sources that no longer matched them. The notes were right. The thing they pointed at had changed.
The folder that filled while the scheduler said ok
A scheduled job reported success for weeks while the folder it was supposed to empty kept filling. The status was a claim the job made about itself.
Before you replace the work, ask who knows the rules already in force
A stakeholder asked for a quick link to see all documents. It sounded like a convenience feature. In that system, it was a way around the access rules.
When all the green checks were answering the wrong question
A security upgrade merged cleanly, tests passed, the service restarted, and health checks passed for four days. The running service was still on the old version. Repair took under an hour. Detection took four days.
Fix the context before you buy the agents
The most common agent failure I see has nothing to do with model quality. An agent gets a task, and the information it needs is spread across chat threads, a wiki, two repositories, and somebody's inbox.
I ended up with two sources of truth. I kept both.
Any organization running AI agents alongside human governance will eventually grow two sources of truth: one where the work happens, one where the decisions live.
Every rule in my agent playbook exists because something broke
Every rule in Chris Luttrell's agent playbook exists because he ran into the failure it prevents.
Containment was only the first step
In August I wrote that a lab slowing its own frontier work was a governance signal worth reading. Two things have happened since, and both say that securing your own environment is what you do before the argument starts.
I built thirty agents and kept six
Building a full AI "company" of specialized agents seemed like the obvious architecture. Running it taught me something different about where agent value actually comes from.
What OpenAI's Cyber-Capability Pause Means for Enterprise AI Governance
OpenAI's cyber-capability pause shows a practical governance lesson for enterprises: AI trust depends on containment, monitoring, authority boundaries, and stop conditions.