The Best AI Agents Look Boring
The prettiest output is usually the one you should not trust. Boring, terse, predictable is a feature, not a limitation.
A founder sent us screenshots of two agent outputs last week and asked which one looked better.
The first one had bold headers, colored callouts, three emojis in the executive summary, and language like "delivering transformative value through strategic execution."
The second one was one line: "Sent 8 emails. Zero replies. Recommending we change the subject line before the next batch. Suggestion attached."
His instinct was that the first one looked more polished. So he had been running with it for three months.
The second style is the one you should trust.
#Why the fancy output is a warning sign
An agent that ships pretty output has spent tokens on presentation instead of thinking. That is not a stylistic complaint. It is a resource-allocation complaint.
Every model has a compute budget per response. The more of that budget you spend on decoration (formatting, tone, superlatives, adjectives), the less is left for the actual work. When you see a two-page executive summary with three graphs and a sidebar for a task that was "send emails and see who replied," someone has traded reasoning for theater.
The fancy output also does something worse: it hides the real signal. In the first example above, buried three paragraphs into "delivering transformative value" was the same fact the boring one led with. Eight emails. Zero replies. The boring version put it in the first six words. The fancy version made you dig for it.
#What boring output looks like
Boring output has a few visible traits. Learn to look for them:
- It leads with the outcome, not the context. "Zero replies" is the first thing you see, not the eighth.
- It uses plain nouns and verbs. "Sent." "Replied." "Failed." Not "engaged," "converted," "optimized."
- It shows numbers as numbers. "8" beats "several." "42 percent" beats "a significant improvement."
- It admits what did not work in the first sentence, not the last.
- It ends with the next action, or a question. No hero closing.
If you had to describe the format in one word, it would be "terse." Not "concise" (which is a nicer word for the same thing but often just means "trimmed marketing copy"). Terse.
#Why owners drift toward fancy
The pattern is not random. Three forces push agents toward polished output:
The default template. Most agent frameworks ship with a "helpful assistant" system prompt that encourages hedging, apologies, and stylistic flourishes. If you never edited it, you got the flourishes.
Owner reaction. If the owner responds to fancy output with "great job!" and to terse output with "is that all?", the agent learns. Feedback shapes future output. If terse gets punished, terse goes away.
Vendor incentive. The company selling you the agent wants the demo to feel magical. Magical demos need adjectives. What sells the agent is not what makes it useful once you have it.
None of these are the agent's fault. All of them can be undone in your setup.
#Retraining an agent to be boring
You do not need to switch models. You need to change the prompt and the reinforcement:
- Add a "response shape" section to your prompt. Something like: "Reply in under 100 words. Lead with the outcome. Use plain nouns. No headings unless the answer is a list. Never open with an adjective."
- Respond to terse output with a plain acknowledgment. "Got it, thanks" is enough. Do not overpraise. That is what trained it to perform.
- When the agent slips back into flourish mode, quote the specific line and ask it to redo just that sentence in the boring style. Do not rewrite the whole message. Point at the exact drift.
Two weeks of this and the output steadies. Not because the model changed. Because the reinforcement loop finally pointed the same direction as the instruction.
#The founder from the top
He switched to the boring style. First week, the change was uncomfortable. Reports felt smaller. Meetings felt shorter.
Second week, he noticed he was catching problems earlier. Because the reports were leading with what did not work, he was seeing it in the first sentence instead of skipping to the summary and missing it.
Third week, he said the boring reports felt like the ones a good employee would have written. That was the tell. Good employees do not send victory laps. They send updates.
#Ready for an agent that just tells you what happened?
If your agent's output feels impressive and you are not sure whether that is a good sign, we can look at it. Usually the fix is the prompt, not the model.
