saved
On the Nature of the Swarm
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Konstantin Dunas argues that inference scaling eventually runs out of context, and compaction, disk offloading, and one-off sub-agents each only partly raise a model’s total context capacity. Persistent sub-agents combine memory with question-time summaries, scaling capacity to roughly one window per agent, and at large scale peer meshes beat trees, echoing Coase and Hayek, though communication cost means swarms are no silver bullet.
ideas
- Total context capacity. Compaction stretches a model past its window, but every summary bets on what will matter later.
- Disk versus sub-agents. Files persist information but not understanding; sub-agents summarize after the question is known but forget everything afterward.
- Hire the sub-agent. Persistent sub-agents keep their context, so agents are parallel over context, not just over time.
- No free lunch. Multi-agent is still inference scaling; the price is lossy, token-costly communication between agents.
- From trees to swarms. Trees force every question through the root, a central planner; meshes let specialists emerge and be reused, with hierarchy added only where it pays.
quotes
“every compaction is a bet. The agent summarizing has to decide now what will matter later.”
“Offloading to disk remembers, but can’t think. Sub-agents think, but can’t remember.”
“Building context is what inference scaling mostly looks like once agents act in the world.”
“The root agent is a central planner.”