Beyond Pipelines: A Survey of the Paradigm Shift toward Model-Native Agentic AI

TL;DR

The paper surveys a paradigm shift in AI from pipeline-based, externally orchestrated agents toward model-native agentic AI in which planning, tool use, and memory are internalized within large language models. It foregrounds reinforcement learning as the engine that enables end-to-end acquisition of agentic capabilities, unifying LLMs, task environments, and learning signals into a single framework. Through a structured review of planning, tool use, and memory, it contrasts traditional pipeline approaches with contemporary model-native methods, highlighting data synthesis and reward design as pivotal levers. The work discusses two key agent applications—Deep Research and GUI Agents—and outlines future directions, including multi-agent collaboration and reflection, with system-vs-model roles evolving toward a scalable AgencyOps ecosystem. Overall, it argues for a coherent progression toward model-native agentic AI that grows intelligence through interaction and experience, rather than merely applying predesigned intelligent behaviors.

Abstract

The rapid evolution of agentic AI marks a new phase in artificial intelligence, where Large Language Models (LLMs) no longer merely respond but act, reason, and adapt. This survey traces the paradigm shift in building agentic AI: from Pipeline-based systems, where planning, tool use, and memory are orchestrated by external logic, to the emerging Model-native paradigm, where these capabilities are internalized within the model's parameters. We first position Reinforcement Learning (RL) as the algorithmic engine enabling this paradigm shift. By reframing learning from imitating static data to outcome-driven exploration, RL underpins a unified solution of LLM + RL + Task across language, vision and embodied domains. Building on this, the survey systematically reviews how each capability -- Planning, Tool use, and Memory -- has evolved from externally scripted modules to end-to-end learned behaviors. Furthermore, it examines how this paradigm shift has reshaped major agent applications, specifically the Deep Research agent emphasizing long-horizon reasoning and the GUI agent emphasizing embodied interaction. We conclude by discussing the continued internalization of agentic capabilities like Multi-agent collaboration and Reflection, alongside the evolving roles of the system and model layers in future agentic AI. Together, these developments outline a coherent trajectory toward model-native agentic AI as an integrated learning and interaction framework, marking the transition from constructing systems that apply intelligence to developing models that grow intelligence through experience.

Figures (12)

• Figure 1: From Pipeline to Model-native: RL-driven evolution of agentic capabilities and applications.
• Figure 2: RL-driven data synthesis as the engine of compute-intelligence conversion.
• Figure 3: Overview of agentic planning methods.
• Figure 4: Overview of agentic tool use methods.
• Figure 5: Overview of long context methods for short-term memory. Method categories include: SW (Sliding Window), CS (Compression & Summarization), RAG (Retrieval-Augmented Generation), PE (Position Encoding extrapolation), LS (Long-sequence Synthesis), and AO (Attention Optimization).

Theorems & Definitions (2)

• Definition 1: Pipeline-based Agentic AI
• Definition 2: Model-native Agentic AI