Automated Specialization of Stateful Agent Systems

In recent years, agentic systems are increasingly being deployed as part of multi-agent workflows, with recent work exploring the automated design of multi-agent workflows through static, monolithic agentic system search or per-query optimizers that rebuild workflows on every input. The former approach lacks adaptability, and the latter disrupts the continuity required for agents to accumulate deep task-expertise through experience. For multi-agent systems to succeed in scaled, dynamic environments, they must not only optimize their collaborative workflows but also evolve the capabilities of the agents themselves. In this work-in-progress, we introduce $\textbf{ASpec}$, a "retain-then-escalate" framework that maintains a core team of persistent, state-trained specialists, resorting to costly architectural mutation only when necessary. This policy's effectiveness relies on a two-stage methodology for agent development: an automated $\textbf{discovery}$ phase to identify promising specialist archetypes, followed by a $\textbf{cultivation}$ phase where selected specialists autonomously deepen their expertise by building private memories from experience gained on a training corpus. Through preliminary experiments, we found that ASpec consistently matches or outperforms state-of-the-art agentic workflow optimization baselines while being an order of magnitude more cost-efficient.