What should make a research agent stop searching?

For a web-research agent, adding another search is straightforward. Deciding when it has enough information seems much harder. Consider a simple task: compare three services and recommend one against specific requirements. The agent finds documentation, searches for alternatives, reads reviews, then keeps finding reasons to investigate further. I’d separate the stopping decision into three checks: Coverage: Has it found evidence for each requirement, including reasons its recommendation might be wrong? Progress: Are recent searches adding useful evidence, or returning variations of information it already has? Budget: Has it reached a limit on time, spending or tool calls? The distinction matters because reaching a budget limit doesn’t mean the research is complete. I’d want the output to say which questions remain unresolved rather than present a confident recommendation simply because execution stopped. Likewise, finding ten pages isn’t necessarily progress if they all repeat the same original source. My preference would be to define the evidence requirements before execution, then return either a supported answer or a clear account of what is still missing. For people building research agents: what stopping rule do you actually use? Have you found a useful way to measure “enough evidence” without making every task require manual review? Disclosure: I’m the founder of Dreamtune Labs.