Large language models are increasingly expected to handle complex, long-horizon real-world tasks whose context demands can grow without bound, yet model context windows remain inherently finite. Recent work explores a paradigm where a main agent decomposes tasks and dispatches subtasks to subagents, which execute and return only summarized results, conserving the main agent's context budget. However, performing this well requires delegation intelligence: the ability to decompose complex tasks, determine when and what to delegate, and integrate returned results into the ongoing workflow. Training data for this capability is scarce in naturally occurring text, and to our knowledge, how to synthesize such data and train models to acquire this capability remains largely unexplored in the open-source community. To bridge this gap, we present a preliminary exploration targeting deep research, a representative long-horizon agent task. Specifically, we design a harness that guides the model toward high-quality task decomposition and delegation, while constraining subagents to return results properly to support the main agent's workflow. The harness-guided trajectories naturally encode correct delegation decisions, which we use as supervised fine-tuning data to internalize delegation intelligence into model weights. Our resulting model, SearchSwarm-30B-A3B, achieves 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, the best results among all models of comparable scale. We will release our harness, model weights, and training data to facilitate future research.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a paper called "SearchSwarm," which explores how AI models can handle long, complicated research tasks without losing track of the big picture.
Sam: That's right. Here's the core problem. When an AI model works on a complex research task — say, answering a question that requires reading dozens of sources — it has to hold all of that information in its "working memory" at once. And just like a desk that gets buried under too many papers, that workspace fills up fast. Researchers call this hitting a "context wall."
Alex: So the model just... runs out of room to think?
Sam: Exactly. And SearchSwarm proposes a solution that's actually quite intuitive. Instead of the model trying to do everything itself, it acts like a manager — breaking the big task into smaller pieces and handing those pieces off to "intern" versions of itself. Each intern works in its own clean, uncluttered environment, then passes a tidy summary back up to the manager.
Alex: So the manager never has to wade through all the raw data directly?
Sam: Right. Think of a CEO who needs a report on a complex topic. They don't personally read every source document. They assign specific questions to assistants, each assistant does focused research, and the CEO only sees the one-page summary. The researchers built a structure they call a "harness" — essentially a set of rules that teaches the model to work this way, delegating instead of hoarding information.
Alex: That's a bit like clearing off your desk so you can focus on the actual decision-making instead of drowning in paperwork.
Sam: That's a good way to put it. The researchers call the skill the model is learning "delegation intelligence." It's not just about offloading work — the model also has to learn *how* to write a good briefing for its sub-agents, so they don't waste effort repeating research that's already been done.
Alex: And these sub-agents — they're actually the same AI model, just given a fresh start?
Sam: Correct. Same model, but handed a clean slate each time. No memory of previous searches, no accumulated clutter. When the sub-agent finishes, it compresses what it found into a concise report. That compression is the key — it forces the system to extract what actually matters rather than carrying around every raw result it ever encountered.
Alex: Does this actually lead to better results in practice?
Sam: The evidence from the paper suggests it does. A mid-sized model using this approach performed as well as — and in some cases better than — much larger models on complex research benchmarks. The argument is that the bottleneck for these tasks isn't raw computing power. It's the ability to stay organized. A smaller model that manages its attention well can outperform a larger one that's simply trying to hold everything in mind at once.
Alex: So the real insight here isn't that the AI is getting smarter in some abstract sense — it's that it's getting better at managing its own limitations?
Sam: Precisely. And the researchers reinforce this through the training process itself. They use a technique called "Supervised Fine-Tuning" — which is essentially showing the model many examples of *ideal* research behavior. Not just correct answers, but correct *processes*: how to break a question down, how to brief a sub-agent clearly, how to verify a result before accepting it. The model learns that organizing its own research is just as important as the information it finds.
Alex: That's a meaningful shift in how we think about AI capability. It's less about raw intelligence and more about good habits of mind.
Sam: And it has a practical implication worth noting. If the limiting factor is organization rather than model size, then you don't necessarily need to build a larger, more expensive system to get better results. You need to teach the system to work more carefully with what it already has. SearchSwarm is one attempt to do exactly that.
Alex: That's a genuinely useful framing — and it makes you wonder how much of what we call "intelligence," in humans or in machines, is really just disciplined attention management. Thanks for listening to ResearchPod.