Working Notes on Agent Systems/Brad Zhang

English indexed field note

Many people think that the biggest problem with Agent is "occasionally getting the wrong answer".

Long post · April 30, 2026 · 1 min read

But the real danger is that the system drifts a little bit in the closed loop, and the more it runs, the more it seems to be correct.

This is actually the systematic drift of the Agent.

Target drift, memory pollution, tool misuse, and evaluation distortion are not essentially single-point bugs, but are all caused by systemic drift.

The most troublesome thing is: all the local nodes look normal, but the whole is unstable.

Therefore, when making an agent, we do not just adjust the prompt, but manage the runtime, memory, tool use, and eval as a system.

I broke this problem down into five levels: concept, principle, consequence, prevention and correction.

Source: Open the original on X