The failure taxonomy

Agents are probabilistic systems doing deterministic work. When they fail they do not crash; they produce confident, well-formatted answers while every metric reads green. Three modes feed one amplifier: tool misuse, step repetition and reasoning-action mismatch, all flowing into error propagation.

The same taxonomy at working resolution: what each mode looks like from the outside, and which harness mechanism catches it. The amplifier is the one to fear; the modes are merely where it starts.

Probabilistic systems, deterministic work

Traditional software is deterministic, which implies same input, same output, and when it does fail it gives off concrete signals. An exception, a log entry, an alert. The bug is found, fixed and a patch is deployed. Thus the failure does not happen again. Agents break this paradigm. They are probabilistic systems doing deterministic work. The same input may succeed nine out of ten times and fail on the tenth. And when agents do fail, they do not crash. They produce confident, well-formatted and authoritative sounding answers. The system appears healthy and every metric reads green. The failure travels downstream undetected.

The modes

In literature there are already attempts to classify types of failures when it comes to agents. The most significant are the following.

Tool misuse happens when the agent calls the right tool with the wrong arguments. This can occur because the agent "feels" it is correct to invent a parameter or entity that does not exist. The operation completes with no results and the agent reports this as a fact, with no errors raised. The user blames data quality.

When an agent repeats the same action without progressing it gets stuck in a loop. This is called step repetition and it causes increased compute consumption while to the telemetry the agent appears to be working. The reality is that the agent is running in circles.

Reasoning-action mismatch occurs when the agent reasons one thing and does another. Its internal chain-of-thought reasons to do one thing and yet it proceeds to execute another. The reasoning can be sound but the execution did not follow it. According to the MAST multi-agent failure taxonomy this represents roughly 14% of agent failures.

The amplifier

Any single failure mode is important and can cause serious issues. However, error propagation is far worse, since it produces false artifacts of success in all downstream steps after the system has already failed. Traditional observability cannot cope with those scenarios because agent failures are semantic and not technical. They look like valid data, not error codes. As such monitoring is blind to them because the system is working. Its answers are wrong.