A new multi-agent research harness beats fixed-scaling approaches on 13 of 15 open-ended optimization tasks by letting an orchestrator adjust parallelism on the fly.
SwarmResearch pairs a "Shepherd Agent" with a population of "Search Agents," each working in its own git branch with local context. The Shepherd holds the global picture and steers exploration across branches. That separation is the point: the researchers argue that single long-running agents fail because they accumulate context until they're locked into one approach, then grind through low-level edits instead of reconsidering the strategy. By isolating each search agent to its own branch and its own context window, SwarmResearch keeps options open longer. On 13 of 15 benchmark tasks, this produced better or comparable results versus leading LLM-guided evolution and multi-agent methods.
The finding pokes at a quiet assumption baked into most multi-agent pipelines: that scaling up agents — whether in series or parallel — is the main lever. SwarmResearch's actual advantage is adaptive parallelism, varying how many agents run and at what search depth based on what the orchestrator learns. That's a different design axis entirely, and it's one most current frameworks don't expose.
The architecture echoes ideas from evolutionary computation — population diversity, branch isolation, a coordinator that selects and steers — applied to LLM agents. Whether it holds up outside coding optimization benchmarks, or at the cost and latency real deployments demand, is the question this paper leaves open.