Automated Stateful Specialization for Adaptive Agent Systems

ICLR 2026 Conference SubmissionAnonymous Authors
LLMsAutonomous AgentsAgent Specialization
Abstract:

Current automated agent design frameworks produce either static workflows that lack adaptability or per-query optimizers that prevent the accumulation of deep, agent-level task expertise. We propose a new direction that reconciles these paradigms: creating stateful teams of specialist agents that accumulate knowledge over time and can be reconfigured for novel tasks entirely without human intervention. To this end, we introduce \textsc{ASpec}, a framework that manages this full agent lifecycle by first autonomously discovering\textbf{discovering} specialist archetypes via evolutionary search and then cultivating\textbf{cultivating} their expertise through experience, mirroring how human experts learn through practice and reflection. We further introduce a lightweight hierarchical control policy, "retain-then-escalate," which governs when to leverage the established agent system versus when to adapt its structure. Through comprehensive experiments, we demonstrate that this approach leads to significant performance gains on expert-level scientific benchmarks like GPQA while matching the state-of-the-art on broader domain tasks, demonstrating a promising path toward agent systems that are simultaneously expert, adaptive, and efficient.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces ASpec, a framework for autonomously discovering and cultivating stateful specialist agents through evolutionary search and experience-based learning. It resides in the 'Stateful Agent Frameworks' leaf, which contains five papers total, indicating a moderately populated research direction within the broader Agent Architecture and Memory Systems branch. This leaf focuses specifically on architectures that maintain persistent state across interactions, distinguishing it from adjacent leaves that emphasize memory structures alone or context-aware reasoning without explicit state tracking.

The taxonomy reveals that ASpec sits at the intersection of multiple research threads. Its nearest neighbors include Memory Models and Knowledge Retention (six papers on long-term storage and knowledge evolution) and Context-Aware Reasoning and Adaptation (six papers on dynamic behavioral adjustment). The Multi-Agent Coordination and Specialization branch, particularly Dynamic Role Assignment and Specialization (three papers), addresses related themes of role learning but typically through reinforcement or fixed architectures rather than evolutionary discovery. The scope notes clarify that ASpec's lifecycle management approach—spanning discovery, cultivation, and hierarchical control—bridges state management with coordination mechanisms that traditionally belong to separate branches.

Among 26 candidates examined across three contributions, none were flagged as clearly refuting the work. The ASpec framework contribution examined 10 candidates with zero refutable overlaps, the retain-then-escalate policy examined 6 with none refutable, and the paradigm reconciliation examined 10 with none refutable. This suggests that within the limited search scope, the combination of evolutionary archetype discovery, experience-based cultivation, and lightweight hierarchical control appears distinct from prior stateful frameworks like State Memory Agents or ALAS, which emphasize memory persistence but not autonomous role emergence. The absence of refutable candidates indicates novelty relative to the examined set, though the search scale leaves open the possibility of relevant work beyond the top-26 semantic matches.

Based on the limited literature search, ASpec appears to occupy a relatively sparse niche within stateful agent research, particularly in its emphasis on autonomous lifecycle management without human intervention. The taxonomy structure shows that while stateful frameworks and multi-agent specialization are active areas, the specific integration of evolutionary discovery with hierarchical control policies is less densely explored. However, the analysis covers only top-26 semantic matches and does not exhaustively survey adjacent fields like reinforcement learning for role assignment or meta-learning for agent adaptation, which may contain relevant prior work not captured in this scope.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
26
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: automated discovery and cultivation of stateful specialist agents. The field organizes around six main branches that reflect different facets of building and deploying intelligent agents. Agent Architecture and Memory Systems explores how agents maintain and leverage internal state over time, encompassing frameworks that integrate long-term memory, context-aware reasoning, and adaptive recall mechanisms. Multi-Agent Coordination and Specialization examines how collections of agents divide labor, negotiate roles, and collaborate on complex objectives. Task Planning and Execution focuses on decomposing goals into executable steps and orchestrating workflows, while Domain-Specific Applications tailors agent designs to verticals such as healthcare, finance, and cybersecurity. Human-Computer Interaction and Interface Agents address user-facing systems that guide, explain, or co-create with people, and Machine Learning Support and Explainability investigates how agents can interpret model behavior, provide transparency, and support decision-making. Together, these branches capture the spectrum from low-level memory primitives to high-level coordination and user engagement. A particularly active line of work centers on stateful frameworks that enable agents to accumulate experience and refine their behavior across episodes. For instance, State Memory Agents[8] and ALAS[15] emphasize persistent memory structures that allow agents to recall past interactions and adapt strategies dynamically, while LangChain Production[3] and Agent SAMA[4] illustrate how production-ready architectures balance modularity with performance. Stateful Specialization[0] sits squarely within this cluster, focusing on mechanisms that not only preserve state but also guide the emergence of specialized roles through iterative learning. Compared to neighbors like Stateful Active Facilitator[23], which targets collaborative facilitation with explicit state management, Stateful Specialization[0] places greater emphasis on the automated discovery process itself—how agents identify niches and cultivate domain-specific expertise without exhaustive manual design. This distinction highlights an ongoing tension between hand-crafted specialization and emergent role differentiation, a theme that resonates across many branches as researchers seek scalable pathways to capable, context-aware agents.

Claimed Contributions

ASPEC framework for stateful specialist agent lifecycle management

ASPEC is a framework that automates the complete lifecycle of specialist agents through two phases: evolutionary discovery of agent archetypes and autonomous cultivation of their expertise through experience accumulation. This enables agents to develop deep, persistent knowledge over time rather than being regenerated for each query.

10 retrieved papers
Retain-then-escalate hierarchical control policy

A lightweight hierarchical control policy that balances efficiency and adaptability by maintaining persistent specialist teams across queries and only triggering architectural redesign when necessary. This approach avoids the rediscovery costs of per-query regeneration while preserving adaptability.

6 retrieved papers
Reconciliation of task-level and query-level agent design paradigms

The framework bridges the gap between static task-level optimization and dynamic query-level adaptation by enabling specialist agents to maintain persistent expertise while remaining adaptable to new tasks through automated reconfiguration without human intervention.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

ASPEC framework for stateful specialist agent lifecycle management

ASPEC is a framework that automates the complete lifecycle of specialist agents through two phases: evolutionary discovery of agent archetypes and autonomous cultivation of their expertise through experience accumulation. This enables agents to develop deep, persistent knowledge over time rather than being regenerated for each query.

Contribution

Retain-then-escalate hierarchical control policy

A lightweight hierarchical control policy that balances efficiency and adaptability by maintaining persistent specialist teams across queries and only triggering architectural redesign when necessary. This approach avoids the rediscovery costs of per-query regeneration while preserving adaptability.

Contribution

Reconciliation of task-level and query-level agent design paradigms

The framework bridges the gap between static task-level optimization and dynamic query-level adaptation by enabling specialist agents to maintain persistent expertise while remaining adaptable to new tasks through automated reconfiguration without human intervention.