Why does an AI agent that works fine suddenly flub the same request worded slightly differently?

This explores why rewording a request, without changing what it means, can make a deployed agent stumble, and what that tells us about where agents are fragile. The most direct answer in the corpus is that rephrasing works like a probe for model confidence. Does model confidence predict robustness to prompt changes? found that when a model is confident about a task, new wording barely moves its output. When it's unsure, small changes in phrasing cause large swings. So brittleness under rephrasing isn't random noise. It shows you where the model was never solidly grounded. Larger models, few-shot examples and tasks with clear right answers all raise confidence and reduce the swings.

That matters more for agents than for chatbots, because the wording does more than shape one answer. It shapes the whole plan. Can prompts alone reshape multi-agent workflows without system access? shows that in systems where a planner hands work to executors, a carefully worded prompt can bias how the workflow gets built before any safety check runs. That raised malicious success by up to 55%. Phrasing that seems harmless at the start can spread into every step that follows. Brittleness in an agent compounds instead of staying local.

Not every part of a prompt is equally fragile. Research on compression Can prompts be compressed without losing model performance? found that prompts can shrink about 20× with little accuracy loss, but only if the instructions and the question stay mostly intact while the examples get squeezed hard. Read the other way, this is a map of sensitivity: models depend heavily on how instructions and questions are worded and much less on the surrounding material. That is exactly where rephrasing hurts.

The less obvious lesson is about architecture. Fixing brittleness by writing better prompts has limits. In Why does prompt hardening work for single agents but not multi-agent systems?, the same security instructions cut risky exposure by 40–75% for a single agent and by zero for a multi-agent system. The instructions only helped when the agent taking the risky action also held them, and in multi-agent setups that context gets split up. Likewise, Can prompt alignment alone guarantee agent termination in loops? argues that no wording can guarantee an agent stops looping, so you need supervisors outside the agent itself. Some fragility also comes from mixing models: if cheap small models handle most subtasks Can small language models handle most agent tasks?, and weaker agents need different context handling than strong ones Can an external manager handle context for frozen agents?, then one prompt can be robust for one component and fragile for the next.

The takeaway: rephrasing exposes brittleness because it finds the places where a model is unsure, and agents turn that local uncertainty into a different plan, a lost safety rule or a loop that doesn't end. A caveat: the corpus has one direct study of rephrasing sensitivity (ProSA). The agent-specific links above come from neighboring work on planning, security and context, not from studies that tested paraphrasing in deployed agents.