Adaptation of Agentic AI

TL;DR

This paper addresses the need for systematic adaptation in agentic AI by proposing a unified taxonomy that spans both agent-focused (A1, A2) and tool-focused (T1, T2) paradigms. It formalizes four paradigms through mathematical notations and illustrates them with concrete, domain-relevant examples, emphasizing how signals from tool execution, agent outputs, and tool performance shape learning. The authors survey representative methods across all four paradigms, compare their trade-offs in cost, data efficiency, generalization, and modularity, and discuss applications in deep research, software development, computer use, and drug discovery. They argue for a co-adaptive future where agents and tools evolve together, leveraging the strengths of each paradigm to build more capable, reliable, and efficient autonomous AI systems while addressing safety and efficiency concerns. The practical takeaway is a roadmap that advocates hybrid architectures—combining robust frozen reasoning cores with modular, trainable tool ecosystems to achieve scalable, continual adaptation in dynamic environments.

Abstract

Cutting-edge agentic AI systems are built on foundation models that can be adapted to plan, reason, and interact with external tools to perform increasingly complex and specialized tasks. As these systems grow in capability and scope, adaptation becomes a central mechanism for improving performance, reliability, and generalization. In this paper, we unify the rapidly expanding research landscape into a systematic framework that spans both agent adaptations and tool adaptations. We further decompose these into tool-execution-signaled and agent-output-signaled forms of agent adaptation, as well as agent-agnostic and agent-supervised forms of tool adaptation. We demonstrate that this framework helps clarify the design space of adaptation strategies in agentic AI, makes their trade-offs explicit, and provides practical guidance for selecting or switching among strategies during system design. We then review the representative approaches in each category, analyze their strengths and limitations, and highlight key open challenges and future opportunities. Overall, this paper aims to offer a conceptual foundation and practical roadmap for researchers and practitioners seeking to build more capable, efficient, and reliable agentic AI systems.

Figures (12)

• Figure 1: Overview of adaptations in agentic AI. Agent: the foundation models serving as orchestration and reasoning modules; Tool: callable components other than the agent model that operate independently, e.g., APIs, ML models, subagents, or memory. We categorize these adaptations into two: agent adaptation (A1 & A2): adapting agent models, and tool adaptation (T1 & T2): adapting tools for agents. See more details in §\ref{['sec:overview']}.
• Figure 2: The structure of this paper.
• Figure 3: Illustration of Four Adaptation Paradigms (A1, A2, T1, and T2). In all the panels, letters highlighted in Red denote the components directly being optimized during adaptation. The red arrows show the sources of adaptation signals. The dotted black lines separate the cases of supervised fine-tuning (SFT) and reinforcement learning (RL).
• Figure 4: Development timeline of A1 methods (agent adaptation with tool-execution result as signal).
• Figure 6: Development timeline of A2 methods (agent adaptation with agent output as signal).