Why AI agents struggle to behave

We want agents that just work. Give them a goal, step back, and let them get the job done. The reality is messier.

New research on deep search agents shows exactly why. These systems operate on long-horizon web interactions. They aren't just predicting the next token; they are executing chains of search, analysis, and decision-making over minutes or hours.

The problem? Fragility.

The researchers found that small reasoning errors early in a chain don't just sit there. They propagate. An agent might misinterpret a search result, which leads it to a suboptimal path, which then poisons the context for every subsequent step. By the time it reaches the end, it’s far off the rails. It isn't a single catastrophic failure; it’s a drift.

This matters for how we build. If you want an agent to actually deliver results, you cannot treat it like a fire-and-forget tool. You need audit trails. You need to see where the path diverged.

When we build agentic systems at Orbari, we focus on modularity. Don't ask one model to solve a long-horizon problem in one go. Break it down. Check the work at every transition. If the agent gets the wrong signal early, you need to catch it before it drifts into the weeds.

The goal isn't a smarter model that magically figures out the end state. The goal is a system that knows when to pause and ask if the current direction still makes sense.

Complexity isn't a feature. It's a risk. Keep the loops tight. Keep the human in the loop for the high-stakes pivots.

***

Source:

SearchAuditor: Auditing and Attributing Failures in Long-Horizon Search Agents | https://arxiv.org/abs/2608.05212