SearchOS treats deep research like an operating system
Everyone's building deep-research agents, and everyone hits the same wall: send an agent off to gather evidence for a long time and it loses the thread. It forgets what it already found, repeats searches, can't tell when it's stuck. Throw multiple agents at it and you get duplicate work, idle workers, and blown search budgets. SearchOS-V1, from Renmin University and Ant Group, treats this like a systems problem instead of a prompting problem.
The move is to stop cramming everything into one giant context window and instead externalize the agent's state, the way an OS externalizes memory. It keeps a Frontier task pool of what's left to do, an Evidence Graph of grounded citations, a Coverage Map of what's answered, and a Failure Memory of search patterns that didn't work. A middleware harness sits between the model and its tools, detecting stalls and enforcing budgets. And it runs pipeline-parallel: as soon as one worker frees up, it gets dispatched to a coverage gap instead of sitting idle.
Results back the framing. On WideSearch it hits 80.3 item-level F1, over four points above the best baseline; on GISA it jumps set F1 by 13.4 points, mostly by recovering answers other agents just miss. The gains come from recall. It stops leaving things on the table.
The bigger idea worth taking away: the bottleneck in multi-agent research isn't the model's intelligence, it's the plumbing around it. Context, coordination, knowing when to quit. Paper at arxiv.org/abs/2607.15257
← Back to all articles
The move is to stop cramming everything into one giant context window and instead externalize the agent's state, the way an OS externalizes memory. It keeps a Frontier task pool of what's left to do, an Evidence Graph of grounded citations, a Coverage Map of what's answered, and a Failure Memory of search patterns that didn't work. A middleware harness sits between the model and its tools, detecting stalls and enforcing budgets. And it runs pipeline-parallel: as soon as one worker frees up, it gets dispatched to a coverage gap instead of sitting idle.
Results back the framing. On WideSearch it hits 80.3 item-level F1, over four points above the best baseline; on GISA it jumps set F1 by 13.4 points, mostly by recovering answers other agents just miss. The gains come from recall. It stops leaving things on the table.
The bigger idea worth taking away: the bottleneck in multi-agent research isn't the model's intelligence, it's the plumbing around it. Context, coordination, knowing when to quit. Paper at arxiv.org/abs/2607.15257
Comments