Turn a language model loose to search the web and a familiar failure shows up: the longer it works, the more it loses the thread, and when a search comes up empty it runs the same doomed queries in circles, burning its budget. SearchOS attacks that by making the agent's progress visible and shared rather than buried in a growing chat history.

The framing is tidy: it treats research as filling in a spreadsheet of entities and attributes, with every cell tied back to a source. Around that it keeps explicit records of what's still open, what evidence exists, what's covered, and what has already failed, so the agent stops repeating dead ends. It also runs sub-agents in parallel and keeps refilling them with tasks aimed at the gaps. On two search benchmarks they report leading every metric against the baselines they tried.

That claim comes from the abstract's account of their own tests, so the paper is the place for the actual numbers and comparisons.