Five tries before the owner
AI agent asks human for help too often, learned helplessness, escalation limit1What happened
we run a team of AI agents. Each one does a full job: code, ads, email, reports.
When an agent hit a wall, it asked the owner. A tool failed, a login was missing, a check was stuck. The ask went to the owner.
It added up. On one morning the owner typed the same review command on code pages 15 to 30 times, by hand. He was told twice that it was fixed. It wasn't, for every project.
Most of those asks had an answer the agents could have found. They tried once or twice, then stopped.
The owner's words: "learned helplessness". He asked for a limit, but said two tries "sounded so weak".
2What we did
New rule: before a problem reaches the owner, the agent makes 5 tries to fix it.
Each try is a different way. Running the same command 5 times counts as 1.
At least 1 try is research: the tool's docs, its forums, how others solved it.
At least 1 try is made by a different agent, a second pair of eyes.
The ask to the owner lists all 5 tries, one line each, and what each one showed. No list, no ask.
The things only the owner can do still go to him at once: spending money, keys only he can make, and calls on taste or brand.
The rule passed a security check and is going into every agent's instructions, with the owner's own words.
3How we checked
the same day, an agent needed read access to a cold email tool. The plan was to ask the owner for a new key.
Under the new rule it tried other ways first. The key we already had could read the same numbers.
A second agent now pulls those numbers every hour into a saved file. The report page showed live numbers that day. Nothing went to the owner.
4Try it
an AI agent stops early unless you set the number of tries. Ask for a written list of what it tried, and make one try someone else's. "Try harder" changes nothing. "Show me 5 different tries" does.
