How do I stop my AI agent from repeating the same mistake?
The short answer
Write the correction down where the agent reads it every time, not only in the conversation where you made it. A correction said once is forgotten when the conversation ends. A correction written as a standing rule applies from then on.
Correcting in the moment is not enough
When an agent makes a mistake, the natural reaction is to tell it there and then. It apologises, fixes the answer, and does better for the rest of that conversation. The next conversation starts without the correction, and the same mistake comes back.
What I do in my own system
In my own system, every correction that should last becomes a written rule, in a file the agents read before they start work. Early on, the system replied to me in English, when every message to me must be in Arabic. The fix was not to repeat the instruction every day. It was one line in the standing rules.
The rules are short and specific: what to do, and an example of the mistake it prevents. “Write clearly” changes nothing. “Reply to me in Arabic, even when I write in English” does.
Keep the rules few and current
Rules pile up. When a new rule contradicts an old one, replace the old one. When the list grows long, rules start to be missed, just as a long conversation loses its oldest parts. Every so often, read the list and remove what no longer applies.
Check that the rule took
After you write a rule, give the agent a task where the old mistake would show. If the mistake comes back, either the rule is unclear or the agent is not reading it.
What this means for you
Each time someone in your team corrects an agent, ask one question: was this written down where the agent will read it next time? If not, the same correction will be needed again tomorrow.