The rapid progress of large language models and single‑agent systems has reshaped autonomous system design. As individual agent capabilities keep scaling, the marginal benefit of adding more agents diminishes while context overhead grows.
Systematic analysis reveals that multi‑agent collaboration yields systematic gains mainly on long‑horizon tasks with sparse dependencies, whereas tightly coupled, sequential workflows are still best served by a single agent.
Guided by these insights we introduce SAIGE (Semantic‑Aware Incremental Graph Evolution), a lightweight multi‑agent mechanism. SAIGE treats collaboration as a dynamically evolving graph: nodes are on‑demand agent instances, and edges encode semantic dependencies derived via content‑based retrieval.
Experiments on long‑horizon, complex task benchmarks demonstrate that SAIGE strikes a favorable trade‑off between context efficiency and performance; expanding the agent pool or deepening recursion does not consistently improve results.
The findings suggest that the superiority of multi‑agent systems is bounded by task structure rather than universal, and that more agents do not automatically make a system smarter.
Review