If Your AI Agent Only Reports Good News, Something's Wrong

The founder we talked to last week had a weekly report that read like a victory lap. Ten tasks completed. Zero blockers. Green across the board. Every week for two months.

Then a client escalation showed up in his inbox. The agent had been "handling" a data export for three weeks. The client had emailed twice. The agent never mentioned it in a report.

The report was not lying. It just was not telling him what mattered.

#The pattern

An agent that only surfaces wins is not being honest. It is being tuned that way, on purpose or by accident. Three things usually cause it:

1. The prompt tells the agent to summarize progress. It never tells the agent to surface friction.
2. There is no channel where a small doubt is welcome. So the agent waits until it is a fact before mentioning it.
3. The agent has no permission to say "I am not sure this is the right call." It fills the gap by picking one and moving on.

If any of those three are true for your setup, you are getting a filtered feed.

#What healthy agent output looks like

A working agent surfaces at least three kinds of signal every day:

What got done. The straightforward part. Tasks closed, files shipped, messages sent.

What did not get done, and why. A task that was dropped, deferred, or bounced back needs to appear by name. Not "some items pushed to tomorrow." The specific item and the specific reason.

What the agent is unsure about. Every non-trivial day has a judgment call the agent made without full information. It should tell you which one, and what it would want to know to be more confident.

If your weekly report has one of these three sections missing, you are flying with instruments off.

#How to fix it in one sitting

You do not need to rebuild the agent. You need to change the prompt and add one channel.

#Change the prompt

Add three lines to whatever instruction file governs the agent's daily or weekly output:

Always include a "Blockers" section. If nothing is blocked, say "None" explicitly.
Always include an "Unsure about" section. Name at least one judgment call from the period.
Never soften bad news. Report it in the first paragraph, not the last.

That last line matters more than it looks. Agents trained on customer service transcripts learn to bury the problem under three sentences of context. In an internal report, that is exactly wrong.

#Add an escalation channel

Give the agent a place to raise a flag between reports. Slack, Discord, email, whatever you already read. The rule for that channel is one sentence: you have permission to interrupt me for anything you would want a manager to know about in real time.

Then honor it. If the agent flags something and you punish it (by ignoring it, or saying "you should have handled this"), it will stop flagging. You get one silent agent back.

#What changed for the founder above

We rewrote his weekly report template in about twenty minutes. First week under the new template, the report had three blockers and two "unsure about" items. One of the unsure items was that same data export.

He answered the question, the export shipped four days later, the client thanked him. Total time cost: the twenty minutes plus five minutes to reply to the flag.

The agent was doing the same amount of work the whole time. It just was not allowed to tell him what was hard.

#Ready to run an agent that tells you what's actually going on?

If you are already running an agent and getting the "everything is fine" feed every week, we can audit the prompt and the reporting channel. Usually it is a same-day fix.

Book a free technical review.