← Back to Paper Radar

From Prompt to Action: A Comprehensive Review of LLM Autonomous Agents

Zainab Rafique, Muhammad Wasim, Mudassar Hussain, Muzammil Hussain, Muhamad Irfan Memon
International Conference on Wireless for Space and Extreme Environments | 2025
This survey reviews LLM-based autonomous agents by classifying single-agent and multi-agent architectures and breaking them into four functional modules: perception, reasoning, planning, and action. It frames their capabilities and failure modes for safety-critical, adversarial, and resource-constrained settings such as space and remote wireless systems.

Problem Statement

LLM agents are being deployed quickly in digital assistants, robotics, and mission planning, yet their reliability in hostile, bandwidth-limited, or safety-critical environments is poorly understood. Existing surveys mostly target general-purpose settings and say little about hallucination, compute limits, and robustness under adversarial or extreme conditions. This paper tries to connect general LLM agent research with the needs of domain-specific autonomous systems.

Key Novelty

  • A taxonomy that organizes LLM agents into single-agent and multi-agent architectures, decomposed into perception, reasoning, planning, and action modules
  • Analysis of LLM capabilities and limitations framed for safety-critical and hostile environments (space, wireless, remote), a framing uncommon in general agent surveys
  • Discussion of prospective standards and metrics for agent dependability and robustness in adversarial settings, plus an agenda of open research challenges

Evaluation Highlights

  • Qualitative literature synthesis only; the abstract reports no new experiments or benchmark numbers
  • Comparative critique of the capabilities LLMs enable (zero-shot generalization, dynamic tool use, human-AI collaboration) against their drawbacks (hallucination, resource constraints, safety risks)

Signal Assessment

3/10 This is a review paper at a niche venue, and its perception-reasoning-planning-action taxonomy closely follows earlier influential surveys (e.g., Wang et al. 2023, Xi et al. 2023). Its main distinct value is the framing around hostile and space environments, not new technical insight.

Methodology

  1. Survey recent literature on LLM-based autonomous agents across domains such as digital assistants, robotics, and mission planning
  2. Classify architectures into single-agent and multi-agent paradigms and decompose them into perception, reasoning, planning, and action modules
  3. Catalog emergent LLM-enabled capabilities and critique real-world limitations, including hallucination, compute/resource limits, and safety
  4. Discuss evaluation standards and metrics for dependability and robustness in hostile environments
  5. Identify open research challenges for deploying stable, efficient, robust agents in wireless, remote, and extreme settings

System Components

Perception Module

Ingests and interprets environmental inputs (text, sensor, or multimodal data) into representations the LLM can reason over

Reasoning Module

Uses LLM inference (e.g., chain-of-thought style reasoning) to interpret context, infer goals, and make decisions

Planning Module

Breaks high-level goals into subtasks and action sequences, possibly with feedback-driven replanning

Action Module

Carries out decisions through tool calls, API chaining, or physical actuation in robotic systems

Single-Agent Architectures

A single LLM-driven agent that handles the full perception-to-action loop

Multi-Agent Architectures

Several LLM agents that collaborate, specialize, or debate to solve complex tasks

Dependability & Robustness Metrics

Proposed evaluation dimensions for agent reliability under adversarial, resource-constrained, or hostile conditions

Results

Metric/Benchmark Baseline This Paper Delta
Architecture coverage General agent surveys Single- and multi-agent taxonomy with 4 functional modules Comparable scope; added extreme-environment lens
Capability analysis Generic capability lists Zero-shot generalization, dynamic tool use, human-AI collaboration Qualitative synthesis
Limitation analysis Hallucination-focused critiques Hallucination, resource limits, safety in hostile domains Broader deployment-oriented framing
Evaluation standards Task-success benchmarks Proposed dependability and robustness metrics for hostile environments Conceptual proposal, no empirical validation

Key Takeaways

  • The perception-reasoning-planning-action decomposition is a useful checklist for designing agent systems and diagnosing failures module by module
  • Deploying agents in remote or bandwidth-limited settings calls for resource-efficient models (smaller or edge LLMs) and safeguards against hallucination before acting autonomously
  • Task-success metrics alone are not enough for safety-critical agents; teams should add robustness, dependability, and adversarial-stress evaluations when deployment conditions are harsh

Abstract

Large Language Models (LLMs) have quickly pushed the frontiers of autonomous agents with advanced reasoning, natural language interaction, and tool chaining in complex worlds. With LLM-based agents arising in domains like digital assistants, autonomous robots, and mission planning, it is more critical than ever to have a deep understanding of their construction, strengths, and weaknesses—especially for safety-critical and adversarial domains like space systems. This article presents an overview of the most recent developments in autonomous agents built with LLMs. We categorize modern architectures, single-agent and multi-agent architectures, and their most prominent functional modules—perception, reasoning, planning, and action. We present new functionality facilitated by LLMs, including zero-shot generalization, dynamic tool use, and human-AI collaboration, and criticize their drawbacks in real-world use, e.g., hallucination, limited resources, and safety. Besides, we discuss future standards and metrics for LLM agents, including how to measure dependability and robustness in hostile environments. Lastly, we present open research challenges highlighting the necessity of stable, efficient, and robust LLM-based agents deployable in wireless, remote, and hostile environments. This survey aims to offer researchers and practitioners a brief overview of the status quo with LLM-based autonomous agents and inspire future work bridging current gaps between general-purpose language intelligence and domain-specific autonomous systems.

Generated automatically from available metadata and abstract on 2026-11-09.