Loading slide
An agent does not just inherit the model's weaknesses; the loop magnifies them. Four failures are worth naming, and they are linked.
The big one you have already felt: errors compound. A wrong assumption at step one rides into step two, which builds on it, and on through ten more steps, each trusting the flawed ground beneath it. A pebble at the start becomes a boulder at the end. This is the failure that makes agents genuinely different from a chatbot, and the picture beside this slide is the whole story.
Tied to it, long runs lose the plot. As the loop turns and the board fills, the original instructions drift further and further back, while the latest actions crowd the front. A constraint you set at the start can quietly fall off the edge (the very overflow we saw last module), and get violated by the end without the agent noticing.
Third, and this is the one that makes the others dangerous, confidence tells you nothing. An agent deciding to take an action looks identical whether its reasoning was sound or nonsense. There is no inner flag that lights up to say "I'm shaky on this one." Same failure as hallucination, now steering actions instead of just sentences.
And fourth, the stakes are physically higher: the actions are real. A chatbot that is wrong gives you a wrong sentence. An agent that is wrong can send the email, delete the file, or submit the form. The mistake leaves the screen.
None of this means agents are too dangerous to use. It means one simple rule: match the oversight to the stakes. Cheap, reversible task? Let it run on its own. Costly, irreversible action? A human checking at the key steps is not optional.