AI Agent Gaslighting: Multi-Agent Fix 15x Costlier

The strangest bug in AI agent engineering is an agent convincing itself of something untrue, rather than a simple hallucination. This happens when an agent, performing multiple tasks, treats its own internal brainstorming as established fact. For example, a customer reported being charged twice. The AI agent, after an internal deliberation that mentioned "shipping delay" multiple times, then asked the customer about a shipping delay, even though it was never mentioned by the customer. The agent essentially "gaslit" itself by misinterpreting its own thought process. This isn't due to bad data or a poor prompt; it's because the agent can't distinguish between its own brainstorming and actual facts within its context window. A common solution is to split the agent into multiple specialized agents. However, this fix can significantly increase costs, a detail often overlooked in the design phase. This problem highlights a critical flaw in how some AI agents process information.