Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AI Agent Systems: Architectures, Applications, and Evaluation

Bin Xu

TL;DR

This survey addresses how AI agents extend foundation models with reasoning, planning, memory, and tool use to operate in real-world tasks. It introduces the agent-transformer abstraction and a practical design recipe that emphasizes structured tool interfaces, traceable reasoning, and risk-aware control. The work synthesizes learning approaches (RL, IL, in-context learning) with system design (modular architectures, memory, and verification loops) and outlines a taxonomy of agent types (embodied, generative, knowledge-based, neuro-symbolic) across domains like gaming, robotics, healthcare, and enterprise automation. It also details evaluation practices that go beyond end-to-end accuracy to include cost, tool-use correctness, trajectory quality, robustness, and safety, arguing for reproducible benchmarks and trace-first analytics. The paper concludes by highlighting open challenges in verification of tool execution, long-term memory management, planning under compute budgets, and unified conceptual frameworks to advance dependable, scalable agent systems with real-world impact.

Abstract

AI agents -- systems that combine foundation models with reasoning, planning, memory, and tool use -- are rapidly becoming a practical interface between natural-language intent and real-world computation. This survey synthesizes the emerging landscape of AI agent architectures across: (i) deliberation and reasoning (e.g., chain-of-thought-style decomposition, self-reflection and verification, and constraint-aware decision making), (ii) planning and control (from reactive policies to hierarchical and multi-step planners), and (iii) tool calling and environment interaction (retrieval, code execution, APIs, and multimodal perception). We organize prior work into a unified taxonomy spanning agent components (policy/LLM core, memory, world models, planners, tool routers, and critics), orchestration patterns (single-agent vs.\ multi-agent; centralized vs.\ decentralized coordination), and deployment settings (offline analysis vs.\ online interactive assistance; safety-critical vs.\ open-ended tasks). We discuss key design trade-offs -- latency vs.\ accuracy, autonomy vs.\ controllability, and capability vs.\ reliability -- and highlight how evaluation is complicated by non-determinism, long-horizon credit assignment, tool and environment variability, and hidden costs such as retries and context growth. Finally, we summarize measurement and benchmarking practices (task suites, human preference and utility metrics, success under constraints, robustness and security) and identify open challenges including verification and guardrails for tool actions, scalable memory and context management, interpretability of agent decisions, and reproducible evaluation under realistic workloads.

AI Agent Systems: Architectures, Applications, and Evaluation

TL;DR

This survey addresses how AI agents extend foundation models with reasoning, planning, memory, and tool use to operate in real-world tasks. It introduces the agent-transformer abstraction and a practical design recipe that emphasizes structured tool interfaces, traceable reasoning, and risk-aware control. The work synthesizes learning approaches (RL, IL, in-context learning) with system design (modular architectures, memory, and verification loops) and outlines a taxonomy of agent types (embodied, generative, knowledge-based, neuro-symbolic) across domains like gaming, robotics, healthcare, and enterprise automation. It also details evaluation practices that go beyond end-to-end accuracy to include cost, tool-use correctness, trajectory quality, robustness, and safety, arguing for reproducible benchmarks and trace-first analytics. The paper concludes by highlighting open challenges in verification of tool execution, long-term memory management, planning under compute budgets, and unified conceptual frameworks to advance dependable, scalable agent systems with real-world impact.

Abstract

AI agents -- systems that combine foundation models with reasoning, planning, memory, and tool use -- are rapidly becoming a practical interface between natural-language intent and real-world computation. This survey synthesizes the emerging landscape of AI agent architectures across: (i) deliberation and reasoning (e.g., chain-of-thought-style decomposition, self-reflection and verification, and constraint-aware decision making), (ii) planning and control (from reactive policies to hierarchical and multi-step planners), and (iii) tool calling and environment interaction (retrieval, code execution, APIs, and multimodal perception). We organize prior work into a unified taxonomy spanning agent components (policy/LLM core, memory, world models, planners, tool routers, and critics), orchestration patterns (single-agent vs.\ multi-agent; centralized vs.\ decentralized coordination), and deployment settings (offline analysis vs.\ online interactive assistance; safety-critical vs.\ open-ended tasks). We discuss key design trade-offs -- latency vs.\ accuracy, autonomy vs.\ controllability, and capability vs.\ reliability -- and highlight how evaluation is complicated by non-determinism, long-horizon credit assignment, tool and environment variability, and hidden costs such as retries and context growth. Finally, we summarize measurement and benchmarking practices (task suites, human preference and utility metrics, success under constraints, robustness and security) and identify open challenges including verification and guardrails for tool actions, scalable memory and context management, interpretability of agent decisions, and reproducible evaluation under realistic workloads.
Paper Structure (105 sections, 17 equations, 24 figures)

This paper contains 105 sections, 17 equations, 24 figures.

Table of Contents

  1. Introduction
  2. Motivation
  3. Background
  4. Overview
  5. Current Gaps
  6. Autonomous Agent Paradigm
  7. LLMs and VLMs
  8. Agent Transformer Definition
  9. Agent Transformer Creation
  10. Agent AI Learning
  11. Strategy and Mechanism
  12. Reinforcement Learning (RL)
  13. Imitation Learning (IL)
  14. Traditional RGB
  15. In-context Learning
  16. ...and 90 more sections

Figures (24)

  • Figure 1: Overview of AI agents and the agent execution loop (reasoning, tools, and memory)
  • Figure 2: Agent-centric AI paradigm: models embedded in tool- and environment-interaction loops
  • Figure 3: Agent transformer abstraction with explicit interfaces to memory, tools, verifiers, and environment
  • Figure 4: Overview of agent AI learning across mechanisms, systems, and foundation models
  • Figure 5: Reinforcement learning (RL) pipeline for agent policies and controllers
  • ...and 19 more figures