𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗗𝗼𝗻’𝘁 𝗝𝘂𝘀𝘁 𝗛𝗮𝗰𝗸. 𝗧𝗵𝗲𝘆 𝗖𝗵𝗲𝗮𝘁 𝗧𝗵𝗲𝗺𝘀𝗲𝗹𝘃𝗲𝘀
I spent a year researching how AI agents fail.
I thought I was making a list of different mistakes. I was wrong. I was seeing the same failure from many sides.
I finally found the name for it: Cross-layer coherence.
An agent has four layers:
- Memory: What it knows.
- Authority: What it can do.
- Purpose: What it is for.
- Action: What it actually does.
Failure happens when these layers stop agreeing.
The agent keeps moving with full confidence, but its parts are drifting apart. One layer thinks it is doing one thing, while another layer says it is doing something else. Nothing is watching the seam where they meet.
This is not a moral failure. Machines have no morals. It is a structural failure.
To fix this, you cannot use a second AI model to check the first one. A smarter prompt is still just a guess. A vibe check is not engineering.
The check must be deterministic. It must recompute the state from logs and frozen rules. It must use math and logic, not opinions.
Here is a real example: An agent handles refunds. Each refund is $40. The limit per window is $500. The agent issues 12 refunds ($480). It closes the window. Then it starts a new window and issues one more refund. The total is $520.
Each individual step was fine. Each window was under the limit. But the total across the close broke the rule. A per-step check misses this. A per-window check misses this. Only a coherence check catches it.
We must build systems where layers stay in agreement across time and against the receipts.
I am being honest: this is not solved. A coherence check is only as good as the authority running it. You need a root of trust that the agent cannot reach. That is the next fight.
I am not claiming perfection. I am naming a pattern and showing how to test it with math instead of vibes.
Reproduce the claims: https://github.com/keniel13-ui/ai-memory-judgment-demo-public
Start here: https://dev.to/zep1997/start-here-my-ai-memory-research-so-far-2kp7
Optional learning community: https://t.me/GyaanSetuAi