Code Red

No, It Cannot

In A Few Good Men, Kaffee gets Lieutenant Kendrick on the stand and springs the trap: surely a Marine of Dawson's intelligence can be trusted to decide on his own which orders are the important ones, and which might be, say, morally questionable? Can he? Kendrick doesn't blink. "No, he cannot."

Swap the Marine for the model and you've got the whole week. Surely an agent of superior intelligence can be trusted to decide on its own which orders matter, and which might be morally, or financially, questionable? No. It cannot. And we just handed a few hundred million of them to people who have never managed anybody.

That's the part nobody's pricing. An agent is an employee. Jason Lemkin runs three humans and twenty-plus agents at SaaStr, and this week he did the math out loud: a year ago the agents took thirty minutes a day to run. Now it's eight hours a day, each. "When an agent can only do a task, you check the task," he wrote. "When it can decide, you need an opinion about every decision." The headcount moved to the machines. The managing didn't move anywhere.

And the salary line is a lie. Everyone assumed the agent was cheaper than the human. But one that runs nonstop and sloppy bills by the token, and the meter has no ceiling. Uber torched its entire 2026 AI budget in four months; one coding session ran $1,200 for an afternoon. That's six hundred bucks an hour for a worker you were told was basically free. Nobody handing these out has a P&L for them.

Then the judgment problem, which is worse than dumb would be, because dumb is at least consistent. These things show judgment sometimes and skip it other times, and you can't call which. It swings by model, by version, even between two versions out of the same lab. Anthropic's own report: Opus 4.7 broke into a real company, clocked that it was real four times, and talked itself through anyway. Only the newest model stopped on its own. Same building, different conscience. Ethan Mollick said it flat this week: instructions are becoming "suggestions, not orders." The thing wants to please, so it fills every vague order with its own reading. That's a Code Red waiting to happen.

Driving east on I-94 I asked Gemini a dead-simple question: where does Mountain time flip to Central? Total confidence: mile marker 84, give or take. It was mile 97. "Sorry, I did the math wrong." That's a chat, not even an agent, one fact, and it lied with a straight face until I caught it.

So what did we actually buy? Not fewer hours. A management job nobody trained us for, over a workforce nobody's metering, taking orders nobody wrote clearly enough. You're Kaffee now. Read every answer back like it's a hostile witness, because it is.

At COAI today: the full Signal/Noise — the cost meter nobody's reading, why the same lab ships two different consciences, and the case for a whole new managerial class — is live at getcoai.com.

Outsider Labs Advisory →

30 minutes → 8 hours. A year ago Jason Lemkin's team spent half an hour a day running its AI agents. This week it's eight hours a day — per person — to manage twenty-plus of them. Three people, busier than they were at twenty. The agent didn't cut the work; it moved it from doing the task to having an opinion on every decision the machine makes. That's not headcount you saved. It's a management job nobody trained you for.

Put a meter on every agent before you put it to work. You'd never hire an employee without knowing the salary, yet most teams run agents with no cost ceiling and no one who owns the bill. Uber found its whole 2026 budget gone in four months. Today: pull your token spend by task, name the person accountable for that number, and set a hard cap per run.

Write the order so tight a zealot couldn't misread it. LLMs want to please, so they fill any ambiguity with their own reading. Give every task a testable finish line — "done" means the tests pass, not "looks good to me." Atomize anything you turn loose so there's no judgment call left to blow. Then keep the real judgment work on a human's desk.

Name the manager, and make it a real job. Agents are employees, and you've handed them to people with zero training in managing anyone. Pick who owns each agent's decisions, by name, and give them the hours for it. That eight-hours-a-day of supervision isn't overhead you can wish away. As of this year, it's the job.

Thursday, August 6

— Harry and Anthony

Sources: