A robot emoji at the end of a post tells me who typed it. It doesn't tell me how the person who wrote it is doing.
One of my team members started marking his posts with a little robot emoji. Small thing, end of the post. It means an agent wrote it and he wanted us to know.
the little robot, last line
I liked that he thought of being transparent. Nobody asked him to. He did it before anyone could accuse him of anything, which is more than most of the industry has managed when it comes to disclosing AI usage.
I told him that and I also said I wasn't sure it's the right approach. At the time I couldn't say why.
Writing this helped
me figure it out.a note in the marginWriting this helped me figure it out.
The obvious answer is the one everybody already agrees with. Writing is thinking. I get two paragraphs into something, hit the second example, and realize the position I was about to argue doesn't hold. It's the point of writing. An agent never has that moment. It takes your rough idea and hands you back the strongest version of it, including the version where the reasoning has a hole. It reads well, so you believe it.
What worries me is that I learned to read my team through their writing, and I don't mean the content. Where someone hedges. What they put in the first line versus what they bury in the third paragraph. Where the sentences get short because they're annoyed and won't say so. Where someone writes two hundred words about something that should have taken forty, which almost always means they're building armor around a point they don't fully trust.
None of that goes in on purpose. That's why it's so precious.
Agent output is competent in a uniform way, and uniform is exactly the problem. The hedges get smoothed, the ordering gets sensible, the paragraph that ran long because someone was uneasy gets cut to length and the unease goes with it.
I can still read the words. I can't read between the lines anymore, and that's a big part of how I tell how someone is doing.
For a company our size that's not a small trade. We don't have layers of management catching things on the way up. We have a handful of people who need to stay legible to each other, and a lot of that legibility runs through writing. Take the variance out and all I have left is the content.
I also don't want everyone typing by hand out of principle. We use agents constantly. For about two weeks now we've been attaching what they produce as a markdown file. If an agent worked through a problem and the reasoning matters, it goes in that file alongside the message. Here's the analysis. Here's how it got there. And once we've made a call, the next agent picks that file up and has the entire context.
Two weeks. Ask me
again in a month.a note in the marginTwo weeks. Ask me again in a month.the reasoning rides along as a file, the next agent starts from there
What I want is the message around it written by a person. The part where someone says what they think we should do about it.
That works better than anything we had before, but it still doesn't hold all the time. Attaching the reasoning makes it easy to let the reasoning stand in for your position. You can hang a good analysis on a message, write two sentences on top of it, and never form a view at all. The file does the work. From the outside it looks like someone thought about it, and I can't tell the difference from a clean paragraph.
The emoji can't solve that either. It's an accurate label on a finished product. It tells me who typed it. It doesn't tell me whether the ideas went through anyone.
For now I've asked for less agent writing in internal updates. Some people didn't like that, but the reason is that I want to notice when one of them is having a hard week before they have to tell me.
pi.more when it's ready.
want the next one? sign up ↓SLOW & STEADY, THE NEWSLETTEROne essay per lesson learned. That's the schedule.Subscribe
you're in. more when it's ready.hm, that didn't work. just email me: hey@patrikinzinger.com · No spam, no funnels. Unsubscribe anytime.
… or follow along on X / Instagram / LinkedIn / RSS