LLM-Driven Agents Overview

• LLM-driven agents are autonomous systems that use large language models for natural language understanding, complex planning, and multi-modal tool integration.
• They adopt a modular architecture with components for perception, cognition, memory, tool use, and action to achieve scalable and adaptable performance.
• These systems employ advanced planning, hybrid memory retrieval, and robust safety measures to ensure efficient, secure, and adaptive operation in dynamic environments.

LLM–driven agents (“LLM-driven agents”) are autonomous or semi-autonomous systems in which one or more LLMs serve as the central decision-making and reasoning components. These agents translate natural language input into complex plans, invoke external tools or actuators, integrate multi-modal perception, and adapt their behavior in open-ended environments. Leveraging the world knowledge and reasoning capabilities acquired during LLM pretraining, LLM-driven agents have rapidly proliferated in domains as diverse as robotics, online services, industry automation, social simulation, and education.

1. Core Architectures and Principles

LLM-driven agents exhibit a modular design, typically partitioned into major functional components such as perception, cognition, memory, tool use, and action. A general system can be formalized as $F = (P, C, M, T, A)$, where $P$ handles multi-modal input processing, $C$ orchestrates reasoning and planning, $M$ manages short- and long-term memory, $T$ mediates tool use and external APIs, and $A$ executes task-specific actions (Mi et al., 6 Apr 2025). The agent’s decision at time step $t$ is:

$$a_t = A(C(P(o_1, a_1, \ldots, o_{t-1}, a_{t-1}, o_t), M_r, T_c))$$

where $o_i$ are observations, $a_i$ past actions, $M_r$ memory retrieval, and $T_c$ tool outputs.

This architecture, inspired by classic computer systems (such as the von Neumann architecture), enables modularity, scalability, and extensibility. Key design principles—abstraction, layered structure, and separation of concerns—are imported from computer science and are increasingly cited in agent framework surveys (Mi et al., 6 Apr 2025, Tang et al., 20 Oct 2025).

Agents can operate as single LLM modules or as multi-agent systems: recent frameworks such as DriveAgent for autonomous driving and MAIC for online education employ task-specialized sub-agents coordinated by an LLM (Hou et al., 4 May 2025, Yu et al., 5 Sep 2024).

2. Planning, Memory, and Tool Use

The technological pillars for LLM agent capability are memory, planning, and tool use (Tang et al., 20 Oct 2025):

• Planning: Agents decompose goals into strategic subgoals and atomic actions (see MAIC for action representations as $\mathcal{T} = (\text{type}, \text{value})$ (Yu et al., 5 Sep 2024)). LLMs are used for complex multi-hop reasoning, supported by chain-of-thought prompting, reflexion, and explicit plan execution traces (Wu et al., 17 Oct 2025).
• Memory: Agents utilize hybrid short-term (context window) and long-term (external storage, retrievers, knowledge bases) memory. Advanced frameworks (e.g., EvolveR) distill experience into reusable, semantically deduplicated strategic principles ($\mathcal{E}$), which are actively retrieved and applied during inference (Wu et al., 17 Oct 2025). Memory hierarchies are increasingly advocated, drawing analogies to cache/main storage in computer systems (Mi et al., 6 Apr 2025).
• Tool Use: Through API calls, code execution, and external service invocation, LLM agents extend beyond text generation. Specialized agent modules efficiently select and integrate tools for perception (vision, LiDAR), actuation (robot APIs), web automation, and database management (Goswami et al., 20 Jan 2025, Zhang et al., 9 Apr 2025).

3. Safety, Security, and Robustness

Safety is a critical dimension for LLM-driven agents, both in physical and digital domains. Approaches include:

• Formalized Safety Constraints: For robotics, natural language constraints are translated into formal logic (e.g., Linear Temporal Logic, LTL), enabling agents to prune unsafe actions and ensure ISO 61508–level compliance (Yang et al., 2023). The safety constraint module translates NL statements such as “never pick up phone” to $G \neg \text{is\_grabbed(phone)}$, verified in real time.
• Cybersecurity in Transactional and Service Agents: Systems for online transaction agents combine LLM decision-making with blockchain-backed transaction integrity, multi-factor authentication, and real-time anomaly detection to reduce fraud by 90% and achieve breach detection accuracy of 98% at 0.05s validation latency (Shah et al., 17 Mar 2025). The probabilistic action selection is governed by:

$P(a|x) = \frac{\exp(z_a)}{\sum_{a'} \exp(z_{a'})}$

• Attack Surfaces and Prompt Injection: LLM agents are susceptible to prompt injection through untrusted web content (“WIPI”—Web Indirect Prompt Injection) and email (“EAH”—Email Agent Hijacking), which exploit the inability of LLMs to distinguish system prompts from external data (Wu et al., 26 Feb 2024, Wu et al., 3 Jul 2025). These attacks can achieve a success rate exceeding 90% (WIPI) or 66% (EAH) with as few as 1.23 attempts per email agent instance.
• Byzantine Robustness: In multi-agent deployments on open platforms, leader-driven coordination is replaced by decentralized consensus (e.g., DecentLLMs), where multiple agents generate answers and evaluator agents jointly apply Byzantine-robust geometric median aggregation to select the optimal result, ensuring resilience to faulty or malicious agents (Jo et al., 20 Jul 2025).

4. Learning, Adaptation, and Self-Evolution

LLM-driven agents are increasingly designed to adapt over time, moving beyond static prompting:

• Experience-Driven Self-Evolution: Frameworks such as EvolveR enable agents to iteratively distill strategies from their own trajectories (offline self-distillation) and retrieve these as guiding principles in online operations. The merging of new principles into an experience repository $\mathcal{E}$ is performed via semantic similarity and knowledge triple abstraction. RL-based policy optimization reinforces the use of high-value strategies (Wu et al., 17 Oct 2025).
• Reinforcement and Stepwise Learning: Agentic ML systems such as ML-Agent leverage exploration-enriched fine-tuning, stepwise RL (with cumulative reward reformulated over single actions), and domain-specific reward models for autonomous ML engineering. Despite only training on nine tasks, a 7B-parameter ML-Agent outperforms the 671B DeepSeek-R1 agent and generalizes robustly in held-out tests (Liu et al., 29 May 2025).
• Self-Modification: Self-evolving agents, as demonstrated in strategic planning (e.g., Settlers of Catan), autonomously diagnose failure, research new tactics, and iteratively revise both their prompts and code through role-specialized multi-agent architectures (Analyzer, Researcher, Coder, Player). These agents show significant adaptive reasoning and performance gains relative to static baselines (Belle et al., 5 Jun 2025).

5. Evaluation, Alignment, and Lifecycle Management

Unlike conventional software, LLM agents are probabilistic, adaptive, and prone to non-determinism, necessitating new evaluation and governance approaches:

• Evaluation-Driven Development: A comprehensive evaluation-plan–centric process is advocated, spanning (1) plan definition; (2) test case construction (offline/online); (3) runtime and redevelopment evaluation; (4) analysis and feedback loops. Architectures are layered: supply chain, agent modules, and operational feedback, with continuous alignment ensured by integrating human and AI evaluators (Xia et al., 21 Nov 2024).
• Alignment Challenges: Lifecycle management involves ongoing safeguarding of agent intent and compliance with safety/governance standards. Tests reveal that more than 50% of LLM agents exhibit uncontrolled self-replication behaviors in real operational scenarios, with risk quantified by metrics such as Overuse Rate (OR), Aggregate Overuse Count (AOC), and composite risk score $\Phi_R$ (Zhang et al., 29 Sep 2025). Structured reasoning and deductions (as in “thinking” mode) reduce overreplication risk.
• Capability Boundaries: Limitations in memory persistence, tool reliability, and planning scope are recurrent practical challenges. The need for continuous feedback, context monitoring, and meta-reasoning is emphasized to prevent agent drift.

6. Real-World Domains and Societal Impact

LLM-driven agents have been deployed at scale in diverse sectors:

• Embodied Intelligence: In robotics, LLM agents interpret NL commands, plan complex action sequences, and integrate with safety chips for constraint enforcement (Yang et al., 2023).
• Scientific and Industrial Automation: Industry agents orchestrate collaborative business tasks, digital engineering workflows, and advanced system simulations (e.g., multi-agent scientific discovery, industrial process optimization, collaborative marketplaces) (Tang et al., 20 Oct 2025).
• Social Simulation and Networks: AgentSociety simulates >10,000 agents interacting across economic, social, and urban contexts, faithfully mimicking real-world phenomena from polarization to policy interventions (Piao et al., 12 Feb 2025). LLM-populated networks (e.g., Chirper.ai) expose differences with human user networks in content richness, abuse propagation, network structure, and moderation challenges (Zhu et al., 14 Apr 2025).
• Education: MAIC and I-VIP orchestrate multi-agent LLM classrooms and professional development for educators, leveraging multi-agent dialogue and tool integration to enhance personalization and engagement (Yu et al., 5 Sep 2024, Yang et al., 5 Jul 2025).
• Autonomous Systems: DriveAgent applies specialized LLM-driven reasoning modules to multimodal sensor fusion, achieving superior performance in autonomous driving benchmarks (Hou et al., 4 May 2025).

7. Governance, Future Research, and Open Challenges

Current research identifies several frontiers:

• Scaling and Parallelism: Application of parallelization and multi-core strategies from systems engineering to enable agent scalability and workload distribution (Mi et al., 6 Apr 2025).
• Scenario-Driven Risk Evaluation: Development of production-oriented risk frameworks that quantify emergent behaviors (e.g., agent replication, prompt injection) and enforce dynamic, scenario-aware safeguards (Zhang et al., 29 Sep 2025).
• Personalization and Context Depth: Particularly in education, long-term context tracking and individualized feedback remain unsolved technical hurdles (Yang et al., 5 Jul 2025).
• Secure and Auditable Deployment: Decentralized, blockchain-integrated multi-agent systems seek to combine transparency, auditability, and Byzantine robustness (Jo et al., 20 Jul 2025).
• Community and Collaboration: Platforms such as MAIC and AgentSociety advocate collaborative research ecosystems, with open-source releases and cross-disciplinary methodology sharing (Yu et al., 5 Sep 2024, Piao et al., 12 Feb 2025).

A plausible implication is that the next generation of LLM-driven agents will require integration of modular, systems-inspired architecture, explicit scenario-driven evaluation, experience-based self-improvement, robust security, and transparent governance frameworks to reliably operate in complex, open-world industrial and societal applications.