Can a better prompt fix an AI agent's habit of searching the same narrow ruts every time?

This explores whether you can fix an AI agent's habit of searching narrowly, or in the same few ruts, just by changing how you prompt it, and whether that fix holds when you switch models. The corpus has no head-to-head test of prompt fixes across model families. It does say a lot about where these search habits come from, and that origin story mostly answers the question: prompting can partly widen what an agent explores, but much of the narrowness is trained into the model, and prompts can only uncover what the model is still able to do.

The most surprising evidence is that the bias often comes from post-training itself. Across 14 pairs of models, raw base models given only a loose system prompt eventually find more distinct solutions than their polished, post-trained versions, as long as you let them make enough attempts Do base models find more solutions than post-trained ones?. Post-training makes models reliable on easy cases but cuts off rare solutions they could once reach. The same pattern shows up in search agents: reinforcement learning collapses the variety of strategies an agent tries down to a few that earn reward, while fine-tuning on varied examples keeps that range Does reinforcement learning squeeze exploration diversity in search agents?. So the search bias differs between models largely because their training differs. A prompt can encourage a model to explore, but it can't bring back options that training removed.

Prompting still does more than you might expect. A single model prompted to play several personas and argue among them can reproduce much of what multi-agent debate achieves Can branching prompts replicate what multi-agent systems do?. That is a real lever against one-track searching: you make the model consider viewpoints it wouldn't reach by default. Researchers studying self-improving agents note that most recent progress comes from this kind of fast, cheap, reversible change to prompts, memory and tools, rather than from retraining weights Do self-improving agents really split into two distinct loops?. That's why people try prompt fixes first.

The more durable answer, though, is to stop depending on the prompt alone. Several notes argue that reliability comes from the harness, the structure built around the model. Memory, reusable skills and fixed protocols take over work the model would otherwise improvise Where does agent reliability actually come from?. Harness improvements found automatically across 51 tasks cut token use by nearly half, and the gains appear to be independent of which model sits inside Can agent harnesses be automatically optimized across many environments?. That independence is what you want when the goal is a fix that works across models. Going further, a meta-agent can design a different multi-agent workflow for each query Can AI systems design unique multi-agent workflows per individual query?. Some tasks need separate, specialized agents that check each other, which one agent can't provide however well it's prompted Do single agents always hit organizational limits?.

The takeaway you may not have expected: when an agent searches narrowly, it often isn't misunderstanding your prompt. It is doing what training rewarded, which was to find the common answer reliably. Prompts can push against that, and a looser prompt on a less heavily tuned model can sometimes beat a carefully engineered prompt on a heavily tuned one. But a related limit holds elsewhere too: prompt instructions alone can't even guarantee that an agent stops looping Can prompt alignment alone guarantee agent termination in loops?. Behavior you need to depend on, across models, generally has to be built into the structure around the model, not requested in the prompt.