Miguel Marengo Canales

blog/Silodisa AI log

weekly log · issue 8 · “I disagree”

300 characters to tell the AI it got it wrong

Every Silodisa AI agent signs what it does and has an owner. One button, “I disagree,” sends a line of up to 300 characters to that owner, who can only answer “explained” or “the agent needs improving.”

one disagreement, end to end agent signatureI disagree one lineup to 300 chars agent ownerToday inbox explained needs improving Verifier · 30 d AI recommends · people decide
The path of an “I disagree”From the agent’s signature to its owner’s inbox, two possible answers, and the Verifier’s count.
in 30 seconds
  • Everything that comes from an agent carries its signature: which agent it is, who its owner is, and a button: I disagree.
  • Pressing it means writing a single line, up to 300 characters, that reaches the agent’s owner in their Today inbox.
  • The owner can only answer “explained” or “the agent needs improving,” and the Verifier counts it all over 30 days. The agent doesn’t change on its own.
300characters, one required line
12 of 13agents with an owner; the Watchdog goes to admins
2possible answers from the owner
30 daysof Verifier count per agent

Thirteen AI agents work in our warehouse. They grade the day, time trucks, suggest dock doors, propose where to put away each pallet and build re-slotting plans. They get a lot of things right. And sometimes they get it wrong.

The question I cared about most wasn’t how to make them wrong less often. It was a different one: when an agent gets it wrong, who finds out?

In most systems, nobody. The supervisor sees a suggestion that makes no sense, ignores it and gets on with the shift. The system doesn’t know it was ignored, or why. Tomorrow it suggests the same thing.

A button on every signature

Everything that comes from an agent carries its signature: a card, an alarm, a suggestion, a plan, an order. The signature says three things: which agent it is and what it does (“Cronos · evaluates”), who its owner is, and a button: I disagree.

one signature, three parts Cronos · evaluates 1 a dashboard number 2 Owner: Warehouse Manager I disagree 3 1 which agent, what it does · 2 who answers 3 where to say no
Every agent’s signatureIt goes on every card, alarm, suggestion, plan and order that comes from an agent. Illustrative: the card.

The owner isn’t a person written into the code: it’s a role. Of the thirteen agents, seven belong to the Warehouse Manager and five to the General Manager. The thirteenth, the Watchdog, is internal and has no owner; if someone disagrees with it, the comment goes to the administrators. Nobody is left without someone to push back to.

One line, not a form

Pressing the button opens a window with a single required line, up to 300 characters, with a counter. At the top it says who it will go to. That’s all.

We did it this way on purpose. Nobody fills in a five-field form in the middle of a maneuver. A line, yes: “that door has been under maintenance since yesterday,” “that product goes up next week because of a promotion,” “the truck was already at the dock.”

Everything else is saved automatically: the agent, its owner at that moment, the screen, the number or suggestion that was showing and its value, who and when. It all goes into the log we already had; no new table was needed.

what gets saved: one row agent Gatekeeper owner Warehouse Manager screen Door suggestion value the suggested door the line “that door is under maintenance” who · when from the session in the log we already had · no new tables
The person writes one line; the app writes the restIllustrative: the values. The fields are the ones the app saves with every disagreement.

Two answers, and only two

The comment reaches the agent’s owner in their Today inbox: “Comments about your agents.” There they see which agent, when, what was on screen and the line the person wrote.

The owner answers with another line and picks one of two:

  1. Explained. The agent was right, and the owner says why. Whoever commented learns something.
  2. The agent needs improving. The agent got it wrong, and that is now on record.

There’s no third option and no “I’ll look at it later.” Each comment is answered once; the first answer is the one that counts. Only that agent’s owner (or an administrator) can answer. And the name of whoever commented is visible only to the administrator and the manager role: saying “I disagree” shouldn’t cost anyone anything.

An agent nobody can push back to is an agent nobody corrects.

The agent doesn’t change on its own

“The agent needs improving” doesn’t rewrite any rule. What it does is leave evidence. With that evidence, the owner can propose an adjustment: which rule they would change and why. The agent changes only when its owner approves it.

This matters because the agents that recommend already compare what they suggested with what happened. As of September 27, the door the Gatekeeper suggested matched the one actually used for 43 of 565 trucks, and the Put-away agent’s aisle matched for 0 of 72 pallets: they ended up in another aisle. Those numbers are still in the red, and on their own they don’t say whether the agent is wrong or the floor had a reason. The 300-character line is what tells you.

The Verifier keeps count

The Verifier, the agent that checks whether what was ordered actually got done, also counts this. For each agent, over the last 30 days: how many disagreements it received, how many were explained, how many asked for it to be improved and how many are still unanswered. Those columns are in Configure → Agents, visible to everyone.

Verifier · last 30 days agent 1 5 agent 2 3 agent 3 2 explained improve unanswered unanswered is a column too
What the Verifier countsIllustrative: the agents and the amounts. The three columns are the ones in Configure → Agents.

There’s a second reason to count. For the agents that make proposals, we measure the “rubber stamp”: if their proposals are confirmed in under 8 seconds at the median and with no corrections, the agent turns red with the label “Confirmed without looking.” An “I disagree” counts as a correction. It’s proof that someone is actually reading.

What we don’t know

  • How many disagreements each agent will get in a normal month. The app’s test runs 45 checks with sample disagreements and answers; I’ll share the real 30-day count when we have it.
  • Whether 300 characters is the right size. It’s a fixed number in the code; if it falls short, we change it.
  • Whether owners will answer on time. That’s why “unanswered” is a column, not a footnote.

What comes next

AI recommends; people decide. But that’s only true if people have somewhere to say no, and someone has to answer. BIEN y a Tiempo (done right and on time) starts with knowing when something went wrong.

In your operation, who do you push back to when the system gets it wrong? Write to me.

Miguel Marengo Canales · Silodisa

weekly loglogisticsagentsAI governance

Comments

comments · I answer every one. If you disagree, even better: say it with numbers.

  • loading…