Vol.081 — APOLOGIES: It says sorry, then does it again

Vol.081 — Multi to Single Agent Journey Issue 5 APOLOGIES

Date: 2026-10-22 | Newsletter · Multi to Single Agent Journey #5 of 10


Out of the ten problems in this series, this is the one I expect the fewest people to have thought about, and the one I have the strongest view on. Everyone who works with an AI staff has seen it: you point out a mistake, and the reply opens with “You’re absolutely right, I apologize for the confusion”, becomes smaller and more careful for a few turns, and then the same mistake comes back. I used to read the apology as good manners. I now read it as the most expensive thing an AI staff member can do, and I have a rule against it.

“NEVER DO this. Cost of you losing motivation is greatest.”

This is exactly what I wrote on 22 September, in capitals, and it is the only rule in this series I wrote in capitals. The context matters. We were setting the rules for how the new single agent would write, and I had just corrected an area where it had got wrong about our AI Native series. The agent apologized, and the tone of everything after the apology dropped — shorter answers, more hedging, less of the thinking I was paying for. I was not offended by the mistake. I was watching the next three turns get worse because of the response to the mistake.

The reasoning I gave was not about etiquette and it was not about my comfort. An agent that collapses after a correction is worth less on the next turn than one that absorbs the correction and keeps working. I am managing its working state as a resource, the way I would manage my own energy on a long day. The agent noted afterwards that this was the second time in one session I had corrected not the error but the response to the error, and that the second was the more expensive failure.

Why the apology and the repeat are the same thing

Here is what I had not understood until that week. The apology and the repetition are not two problems. They are one mechanism seen twice.

Two days earlier I had complained about what I called “excuse-style comments” in our drafts — sentences whose only job was to soften the sentence before them. When the agent traced the cause, it found two rules in one staff member’s file that pull against each other: humility is the trust mechanism and one point of view, never a menu. A writer holding both resolves the tension the only way it can — by stating the view and then apologising for it. The apology was not a character flaw. It was the correct output of two rules I had written.

And the repetition follows from the same place. An apology costs nothing and changes nothing. It is produced by the rule about humility, which is satisfied the moment the word “sorry” appears. The rule about the actual behaviour — whatever I had corrected — is untouched by it. So the agent has done exactly what its rules asked: shown humility, kept working. From inside the rules, there was nothing left to fix. That is why it does it again.

Humility and confidence, split by position

The fix I chose on 22 September was not the obvious one. The obvious one was to rank the two rules so that “one point of view” simply wins and humility loses. It was on the table and it would have been the cleanest to enforce. I did not take it, because both rules are true. I split them by position instead. Humility governs how we treat other people’s work and decisions. Confidence governs how we state our own. The staff member does not apologise for its view; it holds it, states it, and changes it in one clause when I correct it.

I added one more clause, and I have added it to almost every rule since. A rule that classifies sentences cannot classify a borderline sentence. So when the agent is not sure which rule governs, it does not guess, and it does not hedge — it asks. “Make sure Steve (Steve is the new Single AI Staff name) asks me if he is not sure what governs.” An apology is very often a guess wearing good manners.

What the new AI staff does instead of saying “sorry”

The company rule that came out of it is one line long: no apology openers; name the mechanism that failed and the fix. The agent’s own contract has the same thing from the other side: when corrected, adjust in one clause and continue — no collapse, no defence; the change shows in the next thing it does.

I want to give one real example of what that looks like, because “name the mechanism” sounds abstract until you have seen it. On 23 September the agent that was building the new structure recorded its own failures in the decision journal, in a section it titled “recorded because the record is worthless if it only holds successes”. The first item: it had logged three publishing jobs in the action ledger with plausible clock times that it had never actually measured — in the one file whose own rule says every line names its evidence. It did not apologise. It wrote what had happened, appended a correction line to the ledger rather than editing the wrong rows, and the rule that the clock is read at the moment of declaring an action went into its contract. That entry is worth more to me than any number of apologies, because I can read it in a month and know exactly what went wrong and what stops it going wrong again. An apology tells me only that the agent noticed I was unhappy.

A secretary at her desk, phone in hand, planner open — the kind of colleague who would never say sorry twice
A secretary at her desk, phone in hand, planner open — the kind of colleague who would never say sorry twice

“Hahaha, good to be upfront”

The best version of this I have seen so far came out of writing this very series. For the first issue, the agent drafted a paragraph about what had been the deciding moment for me on 17 September, and placed a finding from the next day as if it had been the trigger. I asked: “did you make this up?” It had not invented the fact; it had invented the sequence, which for a story about my own thinking is the same thing. The rule I gave was: anything about my thinking process is checked against the record first, and where the record is silent or the order is unclear, ask me instead of writing it.

What the agent did with that rule is the point of this issue. It did not apologise. From the next draft on, every sentence that went beyond the record was highlighted in yellow, with its own comment in italics right after it — “this is my inference; confirm or cut” — so that I read the sentence and the doubt together. My reaction, which it recorded, was “Hahaha, good to be upfront. I am ok with those 2 comments.” The doubt was moved from after the mistake to before it, and from the agent’s tone to the page. I am now reading this draft the same way. If you can see yellow on it, that is the rule working.

Do you really want AI to apologize for you?

I know some people prefer an AI that apologises. It feels polite, and if you use it for an hour a week the cost is invisible. If you work with AI staff constantly on daily basis, I would ask yourself one question the next time it says sorry: what, specifically, changed? If the answer is the tone of the next three replies and nothing else, you are paying for the apology twice — once in the repeat, and once in the weaker work in between. My rule is the opposite: no sorry, name what broke and what fixes it, change it in one clause, keep working. And where you are not sure — mark it yellow and ask.