GraSP: Graph-Structured Skill Compositions for LLM Agents

The core insight worth sitting with is that this is a retrieval and scoping problem dressed up as an architecture problem: agents fail not because they lack skills but because they can't efficiently ignore irrelevant ones, and DAG structure is the proposed solution to that filtering challenge.

GraSP belongs to a cluster of papers on this site wrestling with the same underlying tension: more capability doesn't automatically produce better behavior. The diversity collapse piece from April 20 ('Diversity Collapse in Multi-Agent LLM Systems') found that scaling up multi-agent teams suppresses useful output rather than expanding it. GraSP hits the same wall from a different angle: scaling up a single agent's skill library produces its own form of collapse. Both papers are essentially arguing that structure and constraint, not raw abundance, are what make capable systems actually perform. The shortest-path generalization paper from April 16 adds a third data point, showing that LLMs fail at longer reasoning horizons partly because they can't manage recursive complexity. Across these three, a pattern is forming around the limits of naive scaling within a single inference context.

The real test is whether GraSP's DAG-based selection holds up as skill libraries grow into the hundreds or thousands of nodes, since the paper's counterintuitive degradation finding implies a threshold effect. If follow-up benchmarks show retrieval latency or graph traversal costs rising superlinearly with library size, the architecture trades one bottleneck for another.