AI agents are a management problem, not a technology problem

For the past year I've run a small fleet of machines in my home that do nothing but serve language models — and a rotating cast of autonomous agents that build software, monitor infrastructure, and maintain their own memory on top of them. It started as a lab. It turned into the clearest management lesson of my career.

Here's the uncomfortable observation: when an agent project fails, the post-mortem almost never says "the model wasn't smart enough." It says the goal was ambiguous, the acceptance criteria didn't exist, nobody verified the output, or the agent quietly wandered outside its lane and nobody noticed for a week. Read that list again and replace "agent" with "new hire." Same list.

Delegation is a skill you already pay for

Every organization already contains people who are excellent at getting good work out of other intelligences — they're called managers, and the good ones do three things relentlessly: they define the outcome before the work starts, they check the work against that definition, and they shrink the blast radius of mistakes while trust is being established.

The scarce skill in the agent era isn't prompting. It's writing an acceptance test before the work begins — and having the discipline to actually run it after.

In my own fleet, the single change that most improved agent output had nothing to do with models or prompts. It was a rule: no change to a live system without stating, in advance, the observable outcome that proves it worked. One change at a time. Verify before stacking the next. Checkpoint so everything is reversible. This is not AI engineering. This is operations management, applied to a workforce that happens to be silicon.

Verification is the new bottleneck

When intelligence gets cheap, the constraint moves. It's no longer "can we produce the work?" — an agent will happily produce ten thousand lines of plausible code overnight. The constraint is "can we tell whether the work is right?" Organizations that treat verification as a first-class function — test harnesses, structured review, honest telemetry about failure rates — compound. Organizations that skim agent output the way they skim a consultant's slide deck accumulate invisible debt at machine speed.

This is why I'm skeptical of the "fire the juniors" narrative. The juniors were never just producing work; they were a training ground for the verification instincts your seniors now run on. Remove the bottom rung and you've optimized this quarter's payroll against next decade's judgment.

What to actually do

The technology will keep improving without your help. The management system around it won't. That's the part you own — and it's the part that decides whether agents make your organization faster or just louder.

This essay is backed by a real system: a smart-routed local inference fleet and an agentic build pipeline I run at home. The build logs live in the journal.