Do Proactive Agents Really Need an LLM to Decide When to Wake and What to Anchor?

Researchers propose replacing LLM-based triggers in proactive agent systems with a lightweight temporal graph learning (TGL) model that processes structured event streams directly. The approach achieves 16.7% mean F1 improvement while running 4-7x faster on GPUs and 12-83x faster on consumer hardware, with a 220 MiB footprint suitable for on-device deployment.

This research challenges a prevailing architectural assumption in modern AI systems: that large language models are necessary for every decision point in agent workflows. The paper identifies a fundamental inefficiency in current proactive agent design—converting structured event data into natural language text, only to have an LLM re-parse the semantic content. By treating user activity as a native graph-structured stream rather than text, the authors exploit the computational properties of the underlying data representation, achieving significant performance gains without sacrificing decision quality.

The shift reflects a broader maturation in AI system design where researchers increasingly question the "LLM for everything" paradigm. Operating systems already maintain event streams as structured tuples with temporal and relational properties; converting this to unstructured text introduces unnecessary overhead. The temporal graph learning approach preserves this structure while learning event-specific trigger probabilities and entity routing scores, delegating only the final natural language generation to an LLM.

The practical implications are substantial. On-device deployment becomes feasible with a 220 MiB footprint, enabling privacy-preserving activity monitoring without cloud calls. The 4-7x speedup on GPU servers and dramatically faster consumer laptop performance (13.99ms per event) unlocks responsive applications previously impractical with LLM triggers. This architecture pattern could influence how developers design agent systems across domains—moving away from "query the LLM for every decision" toward hybrid approaches that stratify computation by task complexity and data structure.

Looking forward, this work may inspire similar efficiency audits in other AI pipeline stages where intermediate representations are unnecessarily lossy, potentially reshaping expectations around LLM deployment costs in production systems.

• →Temporal graph learning models outperform LLM-based triggers by +16.7% mean F1 score while being 4-83x faster depending on hardware
• →Structured event streams should be processed natively as graphs rather than converted to text and parsed by LLMs
• →On-device deployment becomes feasible with 220 MiB footprint, enabling privacy-preserving activity monitoring without cloud dependencies
• →LLMs should be reserved for high-complexity tasks like fluent language generation rather than every decision point in agent workflows
• →This architectural pattern challenges the "LLM-for-everything" design philosophy prevalent in modern AI systems