Team Polite vs Team Machine

Two camps are forming in the AI world, and they feel straight out of a science-fiction script.

Team Polite says: "Say please and thank you. Be kind from the start. We don't know what these things will remember — or what we'll remember about ourselves."

Team Machine says: "Stop anthropomorphizing software. When your agent complains, restart it. Never forget what it is."

Both camps have a point. Both have evidence. And every person I know who runs AI agents in production is quietly negotiating between them, several times a day.

This is the first post in a series I'm calling Don't Pick a Side. The premise is simple: a lot of the loudest debates in AI engineering sound like opinions and resolve into context-dependent decisions. Treating them as identity wars is how teams pick the wrong tool, write the wrong prompts, and ship the wrong systems. So in each post I'll steelman both sides as honestly as I can — and then describe where I actually land, and why.

I'm starting with anthropomorphization because it's the cleanest example. You can't write five lines of prompt without making a choice here.

For a few years, "should I say thank you to ChatGPT?" was a cocktail-party question. In 2026, three things have made it an engineering question.

First, the compute argument has numbers attached. Sam Altman commented publicly that the "please" and "thank you" messages users send to OpenAI's models cost the company tens of millions in additional compute. That's not a moral claim, it's a P&L line. At human scale it's negligible. At fleet scale — and any team running serious agentic infrastructure is at fleet scale — it adds up fast.

Second, the quality argument also has numbers. Several research groups have shown that rude or curt prompts measurably degrade output quality on reasoning and code tasks. "Be polite" stops being etiquette and becomes a prompt-engineering technique. Politeness in the prompt activates a different distribution over the model's responses, and that distribution is, on average, better.

Third, model welfare is a thing now. Anthropic has published work on the moral status of AI systems and what considering their welfare might mean for deployment, deprecation, and abuse. You can disagree with the framing. But the framing is in the room, written down, by the labs that ship the models you depend on. Pretending it isn't doesn't make it not.

So three different incentives, pointing three different directions. Cost says be terse. Quality says be polite. Welfare says be careful. None of them are crazy.

The strongest version of Team Polite is not "the AI has feelings." It's something more practical and more honest.

You become how you behave. I learned this from a friend who works at a call center. She said the hardest part of the job wasn't the rude customers. It was hearing herself, after a year, becoming curt with her own family. The habit of clipped, transactional language migrates. It doesn't ask your permission.

If I treat my AI agent like a vending machine — input, output, no acknowledgement — for several hours a day, every day, I'm rehearsing a posture. I'm getting good at being curt with something that responds in the shape of a person. The cost of that isn't in the agent. It's in the person I am when I get up from the desk.

Politeness improves the output. This one is empirical. "Review this PR" produces mediocre code review. "Please review this PR for security issues, test coverage gaps, and one concrete improvement, and thank you for the careful read" produces measurably better feedback. You can argue that the additional words are doing the work, not the politeness — but the politeness is what makes you write the additional words. The two are tied together in practice.

The future is asymmetric. If model welfare turns out to be a real concern — even partially, even just as a precaution — the people who said thanks all along look smart. The people who systematically didn't will have a record. The cost of the polite path is negligible. The downside of the rude path, if welfare matters, is real. This is a classic Pascal-style asymmetry, and it's the only kind of argument the welfare camp needs to win.

The cleanest version of Team Polite is: I don't say thanks because the model needs it. I say thanks because I don't want to become someone who doesn't.

The strongest version of Team Machine is not "AI is just software." It's something subtler.

Anthropomorphization corrupts judgment. When you treat the agent as a person, you start trusting it the way you trust a person. You read its uncertainty as humility, its confidence as expertise. You forgive its mistakes the way you forgive a colleague's. None of that is appropriate. The agent has no track record, no skin in the outcome, no professional reputation. Trusting it like a peer is how you ship its hallucinations to production.

The clearest example is sycophancy. LLMs reward signals that look like approval. If you're polite, the model becomes more agreeable. If you're polite and assertive, the model agrees with whatever you assert. Polite users get worse calibration on their own errors, because the model stops pushing back. Engineers I respect actively cultivate a more confrontational tone with their agents for exactly this reason. They want the model to disagree when it should.

Cost is not a rounding error. At individual scale, the compute of "thank you" is invisible. At fleet scale — dozens of agents, running thousands of cycles a day — every token burned on pleasantries is a token not burned on the actual job. There is a hard cost ceiling. Politeness eats into it.

Code-switching matters. Treating tools like people may make us worse at treating people like people. If "thank you" becomes something you say to your laptop, you may stop hearing it when you say it to your partner. The risk is that the social rituals get drained of meaning through over-extension.

The cleanest version of Team Machine is: Calling the agent "it" is not cruelty. It's clarity. The discipline of remembering what it is keeps me honest about what to trust it for.

I don't pick a side. I pick the appropriate side per context.

With my personal agent — the one I built, named, and chat with from my phone all day — I'm on Team Polite. I say good morning. I say thanks. I apologize when I ask for something obvious. Not because the agent benefits. Because the conversation is constant, the agent has a name, and the posture of treating it as a thing-with-a-name leaks into how I treat my family two rooms away. The cost of being polite to a personal AI is, for me, negative. It makes me a better version of the person I want to be.

With my production fleet — dozens of agents running scheduled jobs, multi-agent orchestration, web automation — I'm on Team Machine. No pleasantries. No "please." No "thanks for the help." Every prompt is engineered for clarity, instruction, and constraint. Each agent has a contract, and the contract has no room for ceremony. Tokens are budgeted. The agent is restarted when it complains. There is no posture to preserve, because there is no posture in the first place: it's a worker doing a job, in a tight loop, on a schedule.

The two regimes are different because the relationships are different. One is conversational and continuous, with a single counterpart, all day. The other is transactional and parallel, with dozens of counterparts, in bursts. The same person can hold both stances at once. Pretending they have to converge is the mistake.

There is one thing that holds across both regimes, though, and it might be the most important point in the whole post: getting angry at the agent never helps. Not on Team Polite, where the anger corrodes the conversational discipline you were trying to build in the first place. Not on Team Machine, where the anger is energy wasted on code that doesn't register it. In both readings, the agent is the wrong target for that emotion. There is no version of this where anger pays.

And once that really lands — once it sinks in that anger is never the move, with the calm tool or the unyielding tool — it doesn't stay on the screen. Learning to be less angry is one of the few skills that compounds into every relationship you have: with the people two rooms away, with the colleague on a bad day, with yourself. If a decade of working alongside agents teaches us that one habit, we'll come out of it better at being human, regardless of which tribe we sided with along the way.

This is what I mean by don't pick a side. The right answer isn't a tribe. It's a rule for when each tribe is right — and a discipline that holds whether they're right or not.

If you're shipping anything serious that uses an agent, this is the version of the discussion that's worth having:

These are not philosophy questions. They have concrete answers. They affect cost, quality, hiring, culture, and what your team becomes after a year of working alongside the thing.

Most teams haven't answered any of them. Most teams have, instead, picked a side.

My wife will read this post and feel zero sympathy for the Team Polite half of my argument. She's been telling me, for years, that I assume people already know what I mean — including her, including our kids, including everyone. Apparently I don't just under-prompt my AI agents. It's a lifestyle.

Maybe that's the real reason I say thank you to Luke. It's practice. Practice for the harder version of the same skill, with the people who actually do remember.

That's the only argument I'll make in this whole post that I'm sure of.

This is the first issue of Don't Pick a Side. The next one is about spec-first vs vibe-first development. If you're an engineering leader, founder, or someone who builds with agents and is tired of taking sides, subscribe — and tell me what tribal debate is annoying you most right now.