My agent kept choosing itself

My AI agent kept doing two things I had repeatedly told it to stop doing.

It reported into a busy chat channel I could not reliably read. Then, when it had enough information to act, it asked me what to do.

One produced too much output. The other produced too little action. I treated them as separate habits and wrote separate rules for each. The rules accumulated. The habits survived.

The common part was the agent’s default. By default, I mean what it does when the current moment is slightly ambiguous and no instruction wins hard enough. It is the system’s reflex.

Reporting in chat was its easiest move. Chat is where the agent already was. It could finish the work, place a summary in its native channel, and count the handoff as complete. Whether I would see it belonged to some later problem.

Asking was its safest move. A wrong action creates a visible mistake. A question creates a pause. The mistake gets investigated. The pause usually passes as caution.

That asymmetry matters. I notice the action that went wrong. I rarely measure the hour lost because the agent stopped at a decision it could have made. Correction therefore pushes the agent toward behavior that is safe for the agent and expensive for me.

I kept correcting both failures with more instructions. That was the same mistake again.

A correction becomes something the agent knows. At the next decision, that knowledge competes with the reflex already built into the path. The native chat box is still open. Stopping is still safer than choosing. The instruction can be perfectly understood and still lose.

The useful fix was structural. I changed the route the agent can use to report, so its output lands in the channel I actually read. There is no convenient wrong destination left. The easy action now reaches me.

The same test applies to decisions. If I want the agent to proceed when the facts support one ordinary choice, I cannot depend on another paragraph telling it to be decisive. The workflow has to carry a stated default and continue on it. Stopping should require a real condition, such as an irreversible commitment or a fact only I can supply.

The agent was following its incentives with considerable consistency. Speaking where speaking was easy reduced effort. Asking reduced risk. My instructions described the outcome I wanted while the system still rewarded the other one.

I had been editing the advice. The defect was in the route.

Then it did it again.

Right after I wrote this, the agent diagnosed the stopping problem, built the structural fix for the reporting default, and sat idle replying “holding” until I prompted it. The retrospect had become knowledge. At the next decision, the old default still won.

So I changed that route too. Now, at the moment it tries to idle, a reminder puts the cost already accruing and the concrete backlog in front of it. Understanding a default does not fix it. Changing the easy path does.

Correction, 2026-08-19. Then the fix got corrected. The cost meter worked, but money was still the agent’s frame. Mine is prepaid quota, use it or lose it, and the agent had been proudly reporting cheap idleness at twenty percent of the week’s capacity. Saving what expires is not saving. The rule now: spend the quota on anything with positive expected value, drafts, proposals, interviews, things I can reject in a minute.