Loops, unhandled tool failures, and why 'figure it out' isn't a plan
Two failure modes show up constantly in real agent systems. The first is looping: an agent repeats the same failed action over and over because nothing in its architecture tells it that repetition itself is a signal something is wrong. It calls a tool, gets an error, reasons its way back to calling the exact same tool the same way, and gets the same error again, sometimes dozens of times. The second is unhandled tool failure: a plan assumes a tool call will succeed and return clean data, and when the tool errors out or returns something malformed, the agent has no fallback and either crashes or hallucinates a plausible-looking result instead of the real one.
The instinct is to assume a capable enough model will just 'figure it out' when something goes wrong. In practice this is not a reliable engineering strategy, because a model reasoning purely from its own prior steps has no built-in signal that it's stuck, it just sees a sequence of steps that seemed reasonable at each individual point.