A verbosity limit isn’t a style choice. It decides whether the one sentence that actually mattered survives being read at all.
The first time I put a hard sentence limit on an agent’s replies, I did it to stop the noise, not because I had a theory about it. What I found afterwards was more interesting than the noise reduction. An agent that answers in six lines gets read in full. An agent that answers in sixty gets skimmed, and the one line that actually mattered — a wrong assumption it wanted to flag, a risk it wanted me to weigh — sits in paragraph four of six and never reaches me.
Length feels like thoroughness from the inside. From outside, it is a cost the agent has shifted onto whoever has to read it. Every extra sentence is a sentence I have to parse to find out whether it was necessary. Most of the time it was not — it was hedging, or a recap of what I already knew, or a tour through options the agent was never going to recommend anyway. The information density of a sixty-line answer is usually lower than a six-line one, not higher, because nobody edits down to the load-bearing sentence when there is no limit forcing them to.
A cap changes what the agent has to do before it writes anything. Without one, it can think out loud and let the structure emerge as it goes. With one, it has to decide what the reader actually needs before the first word goes down, because there is no room left for the working-out. That is a genuinely different step, and it is the one that produces a defensible answer instead of a narrated one. I would rather an agent spend its effort deciding what matters than spend it writing sentences I then have to do that work on myself.
The failure mode of unbounded replies is not verbosity as an abstract quality — it is a specific, repeatable event. Something important gets stated once, correctly, and then buried under three paragraphs of context that did not need saying. I have missed real warnings this way more than once: a build that quietly failed a test, a dependency version that had drifted, a decision that needed my input before proceeding. Every one of those was in the text. None of them were where I was looking, because where I was looking was the top of a wall of prose that had already trained me to skim.
In practice this means more than telling an agent to be brief, because "be brief" is not something a model reliably self-enforces under its own judgement — its judgement is exactly the thing producing the excess. What works is a hard, numeric cap enforced as a rule, plus a small number of structural constraints that go with it: no preamble, no recap of the request, no closing summary, decisions presented as a short list of options rather than paragraphs of argument. Those constraints do the same job from a different angle — they remove the places verbosity likes to hide.
This matters more, not less, as the number of agents I have running at once goes up. One verbose agent is an annoyance. Six verbose agents running in parallel, each reporting back in its own wall of text, is an attention tax that scales with headcount instead of with information. Terseness stops being a style preference at that point and becomes a load-bearing part of the system — the thing that decides whether I actually see the one line I needed, or whether it scrolls past while I am three agents deep into the queue.
None of this means an agent should never explain itself. It means the explanation is a thing I ask for, not a thing that arrives by default, padded around the answer I actually needed. The cap is not there to make replies shorter for its own sake. It is there because the one sentence that matters is only useful if it survives being read.