Where they fail

Agents amplify both the strengths and the weaknesses of the underlying model.

Errors compound. A wrong assumption at step one can propagate through ten subsequent steps, each building on the flawed foundation. By the end, the agent may be confidently doing the wrong thing entirely.

Long chains lose coherence. As the agent loop runs and the context grows, the original task description gets further away from the most recent actions. Constraints stated at the beginning may be violated by the end.

Confidence doesn't signal correctness. The agent "deciding" to take an action looks the same whether the reasoning is sound or not. There's no internal flag that says "I'm less sure about this step."

Real-world consequences. A chatbot that hallucinates gives a wrong answer. An agent that hallucinates may send an email, delete a file, or submit a form. The blast radius is larger.

None of this means agents are too dangerous to use. It means the level of oversight should match the stakes. For low-consequence tasks with recoverable errors, agents are fine to run autonomously. For consequential, irreversible actions, human review at key steps is not optional.