The depth problem with agentic research

Moe Khalil Jul 20, 2026 1 Share I started working on AI research back in 2023, when Perplexity was still a seed-stage startup. Over time, I’ve seen it get better in two ways: more accurate, and deeper. At some point, I believe it hit a wall on both. In this article, I’m going to make my argument on what I believe is the barrier on the depth side. I’ll add a disclaimer that I run a deep research company, where we let you control the depth via a very simple metric, but more on that later.

Agents are trained to complete tasks. That makes sense for coding, which mostly consists of tasks with a very clear finish line. But research is different. You can ask three different people the same research question and have one do a quick google search, one spend a couple of hours deep diving, and one spend weeks going deeper than anyone has before. Unfortunately, current agents are trained to do the bare minimum to complete the task. This is great for coding, but a huge limitation on open ended research.

Research agents need a depth dial. As a user, I should be able to control how much research work goes into a specific question. Every agentic task already has a cost, which is a direct result of effort and model pricing. For the same model, and with proper context compression, cost scales roughly linearly with the amount of work done.

Right now, the amount of work done is decided by a combination of the “effort level,” the model’s biases, and how many “work very hard”s are in your prompt. However, given that linear relationship, it stands to reason that if we can control the cost as a primitive, we can control how much effort goes into a task.

This allows us to break through the depth wall for tasks with a vague stopping point.

The obvious problem is that more time does not guarantee better research. An agent can spend hours repeating the same searches or following bad leads.

A larger budget has to produce more value. It should find evidence the smaller run missed, investigate disagreements between sources, and expose gaps that remain unresolved.

Making this true is possible (barring information saturation in the search space), but requires a proper harness and fine-tuned prompting built to avoid biases or infinite loops within the research budget

As intelligence becomes a commodity, the edge shifts to who has better information, not who has the better model. Two agents running the same model will reach different conclusions if one spent an hour digging and the other spent five minutes. That gap compounds every time the agent acts on what it found.

Deeper research is one of the biggest advantages you can buy.

I run Webhound, and we built the product around this idea, so I am biased.

You give Webhound a prompt and a budget. It uses that budget to conduct the research, then returns the result along with the working documents and sources behind it.

You can try a free $5 run (or about 75 minutes of research depth) here: https://webhound.ai