A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

Thorsten Händler

Ferdinand Porsche Mobile University of Applied Sciences (FERNFH), Austria

Keywords:

Taxonomy, Autonomous Agents, Multi-Agent Collaboration, Large Language Models (LLMs), AI System

Classiﬁcation, Alignment, Software Architecture, Architectural Viewpoints, Software-Design Rationale,

Context Interaction, Artiﬁcial Intelligence, Domain-Ontology Diagram, Feature Diagram, Radar Chart.

Abstract:

Large language models (LLMs) have revolutionized the ﬁeld of artiﬁcial intelligence, endowing it with so-

phisticated language understanding and generation capabilities. However, when faced with more complex and

interconnected tasks that demand a profound and iterative thought process, LLMs reveal their inherent lim-

itations. Autonomous LLM-powered multi-agent systems represent a strategic response to these challenges.

While these architectures hold promising potential in amplifying AI capabilities, striking the right balance be-

tween different levels of autonomy and alignment remains the crucial challenge for their effective operation.

This paper proposes a comprehensive multi-dimensional taxonomy, engineered to analyze how autonomous

LLM-powered multi-agent systems balance the dynamic interplay between autonomy and alignment across

various aspects inherent to architectural viewpoints such as goal-driven task management, agent composition,

multi-agent collaboration, and context interaction. Our taxonomy aims to empower researchers, engineers,

and AI practitioners to systematically analyze the architectural dynamics and balancing strategies employed

by these increasingly prevalent AI systems. The exploratory taxonomic classiﬁcation of selected representa-

tive LLM-powered multi-agent systems illustrates its practical utility and reveals potential for future research

and development. An extended version of this paper is available on arXiv (Händler, 2023).

1 INTRODUCTION

In recent years, the emergence and the technological

feasibility of large language models (LLMs) have rev-

olutionized the ﬁeld of artiﬁcial intelligence (Brown

et al., 2020; Ouyang et al., 2022; Thoppilan et al.,

2022; Chowdhery et al., 2022; Zhang et al., 2022).

Pre-trained on vast amounts of text data, these models

have catalyzed signiﬁcant advancements by enabling

sophisticated language understanding and generation

capabilities, opening doors to a broad range of appli-

cations (Bommasani et al., 2021; Bubeck et al., 2023;

Kaddour et al., 2023). Yet, despite their remarkable

capabilities, LLMs also have inherent limitations.

While LLMs excel at generating outputs based

on patterns identiﬁed in their training data, they lack

a genuine understanding of the real world. Conse-

quently, their outputs might seem plausible on the sur-

face, but can be factually incorrect or even halluci-

nated (Maynez et al., 2020; Ji et al., 2023). Moreover,

despite their proﬁciency in handling vast amounts of

textual information and their rapid processing and

https://orcid.org/0000-0002-0589-204X

pattern recognition capabilities, LLMs struggle with

maintaining consistent logic across extended chains

of reasoning. This deﬁciency hinders their ability to

engage in a deliberate, in-depth, and iterative thought

process (aka slow thinking) (Sloman, 1996; Kahne-

man, 2011; Fabiano et al., 2023; Lin et al., 2023). As

a result, LLMs encounter difﬁculties when it comes

to handling more complex and interconnected tasks

(Kojima et al., 2022; Wei et al., 2022).

These limitations of individual LLMs have led

to the exploration of more sophisticated and ﬂexi-

ble AI architectures including multi-agent systems

that aim at accomplishing complex tasks, goals, or

problems with the cognitive synergy of multiple au-

tonomous LLM-powered agents (Torantulino et al.,

2023; Nakajima, 2023; TransformerOptimus et al.,

2023; Park et al., 2023; Shen et al., 2023; Li et al.,

2023; Shrestha et al., 2023; Hong et al., 2023). Such

systems tackle user-prompted goals by employing a

divide & conquer strategy, by breaking them down

into smaller manageable tasks. These tasks are then

assigned to specialized agents, each equipped with

a dedicated role and the reasoning capabilities of an

LLM, as well as further competencies by utilizing

Händler, T.

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures.

DOI: 10.5220/0012239100003598

In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 3: KMIS, pages 85-98

ISBN: 978-989-758-671-2; ISSN: 2184-3228

contextual resources like data sets, tools, or further

foundation models. Taking a cue from Minsky’s so-

ciety of mind theory (Minsky, 1988), the key to the

systems’ problem-solving capability lies in orches-

trating the iterative collaboration and mutual feedback

between these more or less ’mindless’ agents during

task execution and result synthesis.

One of the central challenges for the effective

operation of LLM-powered multi-agent architectures

(as with many AI systems) lies in ﬁnding the opti-

mal balance between autonomy and alignment (Yud-

kowsky, 2016; Bostrom, 2017; Russell, 2022; Wolf

et al., 2023; Hong et al., 2023). On the one hand,

the systems should be aligned to the goals and in-

tentions of human users; on the other hand, the sys-

tems should accomplish the user-prompted goal in a

self-organizing manner. However, a system with high

autonomy may handle complex tasks efﬁciently, but

risks straying from its intended purpose if not sufﬁ-

ciently aligned, resulting in unexpected consequences

and uncontrollable side effects. Conversely, a highly

aligned system may adhere closely to its intended pur-

pose but may lack the ﬂexibility and initiative to re-

spond adequately to novel situations. Current systems

exhibit diverse approaches and mechanisms to inter-

twine these cross-cutting concerns (Kiczales et al.,

1997) throughout their architectural infrastructure and

dynamics.

However, existing taxonomies and analysis frame-

works for autonomous systems and multi-agent sys-

tems (see Section 2.1) fall short in providing means

to categorize and understand these challenges and

involved architectural complexities posed by LLM-

powered multi-agent systems.

This paper

aims to bridge this gap by introduc-

ing a systematic approach in terms of a comprehen-

sive multi-dimensional taxonomy. This taxonomy is

engineered to analyze and classify how autonomous

LLM-powered multi-agent systems balance the inter-

play between autonomy and alignment across their

system architectures.

A simpliﬁed overview of the dimensions and lev-

els applied in our taxonomy is represented by the

cuboid shown in Fig. 1. First, the synergy be-

tween autonomy and alignment manifests as a two-

dimensional matrix with multiple hierarchical lev-

els. This matrix captures a spectrum of nine dis-

tinct system conﬁgurations, ranging from systems

that strictly adhere to predeﬁned mechanisms (rule-

driven automation, L0/L0) to those that dynamically

adapt in real-time, guided by evolving conditions and

user feedback (user-responsive autonomy, L2/L2).

An extended paper version (Händler, 2023) is avail-

able at https://doi.org/10.48550/arXiv.2310.03659.

Alignment

static / adaptive / self-organizing

integrated

Architectural

Viewpoints

Autonomy

user-

guided

real-time

responsive

goal-driven task management

agent composition

multi-agent collaboration

context interaction

Figure 1: A simpliﬁed representation of the proposed multi-

dimensional taxonomy for autonomous LLM-powered

multi-agent systems. The x-axis represents the level of au-

tonomy, the y-axis the level of alignment, and the z-axis the

four applied architectural viewpoints.

Second, these conﬁguration options are applied to

multiple distinct architectural viewpoints (Kruchten,

1995), such as the system’s functionality (goal-driven

task management), its internal structure (agent com-

position), its dynamic interactions (multi-agent col-

laboration) as well as the involvement of contextual

resources such as tools and data (context interaction).

Stemming from these four viewpoints, we have dis-

cerned 12 architectural aspects, each with distinct au-

tonomy and alignment levels. This granularity facili-

tates a nuanced analysis of the system’s architectural

dynamics resulting from the interplay between auton-

omy and alignment across the system architecture,

laying the foundations for further analysis and rea-

soning about design decisions. The contributions of

this paper can be categorized as follows:

1. Multi-Dimensional Taxonomy. We introduce a

comprehensive multi-dimensional taxonomy tai-

lored to analyze and understand how autonomous

LLM-powered multi-agent architectures balance

the dynamic interplay between autonomy and

alignment across different architectural aspects.

For this purpose, our taxonomy provides hierar-

chical levels for both autonomy and alignment,

which are applied to distinct architectural view-

points and aspects, thus incorporating a third di-

mension.

2. Taxonomic Classiﬁcation of Selected Systems.

We demonstrate the utility of our taxonomy

by classifying a selection of seven autonomous

LLM-powered multi-agent systems, which pro-

vides insights into the architectural dynamics of

the analyzed systems and identiﬁes challenges

and development potentials. The taxonomic ap-

plication also serves as a ﬁrst empirical validation.

Structure of the Paper. The remainder of this pa-

per is structured as follows. Section 2 gives a short

overview of related background. In Section 3, we in-

troduce our multi-dimensional taxonomy, incorporat-

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems

ing speciﬁcations of autonomy and alignment levels

and their application to the system architecture. By

analyzing selected multi-agent systems, Section 4 il-

lustrates the utility of our taxonomy. Finally, Section

5 discusses key insights and concludes the paper.

2 BACKGROUND

2.1 Related Work

Existing Taxonomies. Taxonomies represent struc-

tured classiﬁcation schemes employed to categorize

objects in a hierarchical manner according to spe-

ciﬁc criteria. They ﬁnd applications in a wide range

of disciplines and domains. The ﬁeld of agent sys-

tems spans a variety of conﬁgurations and operational

structures, with some systems operating as individ-

ual entities and others involving multiple interacting

agents.

• Taxonomies for Autonomous Systems mainly

categorize systems based on the level and type

of autonomy, intelligence, learning capabilities,

and ability to interact with their environment

(Wooldridge and Jennings, 1995; Brustoloni,

1991; Maes, 1995; Franklin and Graesser, 1996;

Tosic and Agha, 2004).

• Taxonomies for Multi-Agent Systems extend

beyond the conﬁnes of individual agent charac-

teristics, integrating the dynamics of interactions

and collaborations among multiple agents (Bird,

1993; Dudek et al., 1996; Van Dyke Parunak

et al., 2004; Moya and Tolk, 2007).

While these taxonomies have contributed signiﬁ-

cantly to our understanding of autonomous agents

and multi-agent systems, they were developed prior

to the advent of large language models (LLMs), and

thus fall short in providing means to categorize and

understand the speciﬁc challenges and involved ar-

chitectural complexities posed by autonomous LLM-

powered multi-agent systems. Moreover, while the

concepts of autonomy and alignment are often dis-

cussed in AI literature (Narendra and Annaswamy,

2012; Russell, 2019) and also the system’s architec-

ture plays a fundamental role in software engineer-

ing (Bass et al., 2003), none of these existing tax-

onomies has so far applied a systematic approach to

either investigate architectural aspects or combine the

concepts of autonomy and alignment.

Current LLM-powered Multi-Agent Systems. In

response to limitations of large language models

(LLMs) handling task complexity (Kaddour et al.,

2023), autonomous multi-agent systems utilizing the

reasoning abilities of LLMs have emerged (see Sec-

tion 2.2). Currently, several projects are established

that aim at realizing such autonomous AI architec-

tures for accomplishing complex tasks based on mul-

tiple interacting agents and powered by large lan-

guage models (LLMs). Exemplary but represen-

tative autonomous multi-agent systems are AUTO-

GPT (Torantulino et al., 2023), BABYAGI (Naka-

jima, 2023), SUPERAGI (TransformerOptimus et al.,

2023), HUGGINGGPT (Shen et al., 2023), CAMEL

(Li et al., 2023), AGENTGPT (Shrestha et al., 2023)

and METAGPT (Hong et al., 2023). A recent sur-

vey is provided by (Wang et al., 2023), which focuses

on investigating and comparing the agents’ charac-

teristics and capabilities in terms of proﬁle gener-

ation, memory operations and structures, planning,

tool integration and learning strategies. Complement-

ing this, another recent survey (Xi et al., 2023) of-

fers an overview of existing approaches, contextualiz-

ing them with foundational technical, methodical, and

conceptual paradigms. However, as we dive into the

speciﬁcs of current autonomous LLM-powered multi-

agent systems, striking the right balance between au-

tonomy and alignment emerges as a central challenge.

Given the exploratory state of the ﬁeld, current sys-

tems exhibit a wide range of architectures, each with

its unique strategy for balancing autonomy and align-

ment dispersed across various architectural compo-

nents and mechanisms. The complexity of these sys-

tems underscores the importance of a taxonomy that

can provide a structured understanding and compari-

son of these systems.

2.2 System Characteristics

In the following, we shortly outline the main archi-

tectural characteristics of autonomous LLM-powered

multi-agent systems, as illustrated in Fig. 2.

Alignment Techniques

Memory

Role

Prompt

collaborate

executes

interacts

with

develops

synthesis

Agent

Human

User

Large Language

Model (LLM)

Agent-Interaction Layer

User Interface

Response

Prompt

Preferences

Expert Tools

Data

Foundation Models

Context

System

Architect

specifies

break

down

Goal

Task

Result

...

Figure 2: Overview of the primary characteristics of au-

tonomous LLM-powered multi-agent systems, enhanced by

contextual resources like tools and data.

G Goal-Driven Task Management. Autonomous

LLM-powered multi-agent systems are designed

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

to accomplish user-prompted goals or complex

tasks. For this purpose, the system employs an in-

teractive and multi-perspective strategy, by break-

ing down complex tasks into smaller, manageable

tasks, which are subsequently distributed among

various agents, each equipped with speciﬁc com-

petencies. A crucial aspect of this divide & con-

quer strategy lies in the effective orchestration and

the subsequent synthesis of partial results.

A LLM-Powered Intelligent Agents. Intelligent

agents structure the system as the foundational

components. Each agent is endowed with a

unique set of competencies, which include a

clearly deﬁned role and an individual memory.

The backbone of their reasoning and interpreta-

tive capabilities is rooted in the incorporation of

large language models (LLMs). This enables the

agents not only to reﬂect upon the tasks or to plan

and process the assigned tasks efﬁciently, but also

to access and utilize contextual resources, as well

as to communicate with other agents.

M Multi-Agent Collaboration. The interaction

layer provides the workspace for a network of

LLM-powered agents. While executing the as-

signed tasks, these agents collaborate with each

other via prompt-driven message exchanges to

delegate responsibilities, seek assistance, or eval-

uate task results. Key to the agents’ collabora-

tion is to effectively combine the strengths of each

agent (cognitive synergy). The power of these sys-

tems emerges from the coordinated efforts of the

collective (society of mind (Minsky, 1988)).

C Context Interaction. Some tasks require the uti-

lization of contextual resources, such as expert

tools, data, further specialized foundation models,

or other applications. These resources extend the

agents’ ability to gather environmental informa-

tion, create or modify artefacts, or initiate external

processes, thus enables the agents to effectively

execute complex tasks.

B Balancing Autonomy and Alignment. The dy-

namics of LLM-powered multi-agent systems are

characterized by a complex interplay between au-

tonomy and alignment. As captured in Fig. 3,

this complexity can be traced back to the tri-

adic interplay and inherent tensions among hu-

man users, LLM-powered agents, and governing

mechanisms or rules integrated into the system.

Alignment, in this context, ensures that the sys-

tem’s actions are in sync with human intentions

and values. On the other side of the spectrum,

autonomy denotes the agents’ inherent capacity

for self-organized strategy and operation, allow-

Human

User

LLM-powered

Agents

Automation / Autonomy

Automation / Customization

Supervision / Autonomy

Alignment

Collaboration

Rules &

Mechanisms

System Operation

Figure 3: Triadic interplay and dynamic tensions between

the decision-making entities in LLM-powered multi-agent

systems.

ing them to function independent of predeﬁned

rules and mechanism and without human super-

vision. Moreover, in systems steered by user-

prompted goals, it becomes pivotal to distinct

between generic alignment aspects, in terms of

mechanisms predeﬁned by system architects, and

user-speciﬁc preferences customized by system

users. However, from an architectural perspective,

autonomy and alignment transform into cross-

cutting concerns (Kiczales et al., 1997). They tra-

verse components and mechanisms across the ar-

chitectural infrastructure and dynamics. Achiev-

ing a balanced conﬁguration of autonomy and

alignment is a crucial challenge, which directly

impacts the system’s efﬁciency and effectiveness.

For a comprehensive discussion of related work as

well as for a detailed speciﬁcation and analysis of rel-

evant architectural concepts, please refer to extended

paper version (Händler, 2023).

3 MULTI-DIMENSIONAL

TAXONOMY

In this section, we introduce the system of our

multi-dimensional taxonomy, engineered to method-

ically analyze the interplay between autonomy and

alignment across architectures of autonomous LLM-

powered multi-agent systems. The taxonomy weaves

three crucial dimensions, i.e. hierarchical levels of au-

tonomy and alignment as well as architectural view-

points. Together, they form a three-dimensional ma-

trix for classifying system conﬁgurations (see Fig. 1).

Section 3.1 delves into the complexities of the

interplay between autonomy and alignment. Subse-

quently, Section 3.2, underscores the importance of

incorporating architectural viewpoints into the taxo-

nomic system. Rather than applying the autonomy-

alignment matrix ﬂatly, we propose analyzing each

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems

architectural viewpoint as well as further inherent

architectural characteristics individually. Such a

viewpoint-focused approach allows for a deeper and

more nuanced understanding of the systems, reﬂect-

ing the architectural dynamics and complexities. Fi-

nally, in Section 3.3, we unify these components,

mapping the autonomy-alignment dimensions and

levels onto aspects inherent to the viewpoints.

3.1 Interplay Between Autonomy and

Alignment

Autonomy and alignment, as interdependent and in-

terplaying concepts, have their roots in management

sciences and organizational behavior, playing inte-

gral roles in the ways teams and systems function

(Mintzberg, 1989; O’reilly Iii and Tushman, 2008).

In these ﬁelds, autonomy typically refers to the de-

gree of discretion employees or teams possess over

their tasks, while alignment denotes the degree to

which these tasks correspond to the organization’s

overall objectives. In the ﬁeld of AI, the interplay be-

tween autonomy and alignment remains pivotal (Rus-

sell et al., 2015; Bostrom, 2017). AI systems, by na-

ture, operate with varying degrees of independence

and are often designed to accomplish complex tasks

that are potentially beyond human capabilities. How-

ever, uncontrolled autonomy can pose risks. If the

goals of an AI system deviate from those of its hu-

man supervisors, it could lead to unforeseen conse-

quences or uncontrollable side effects. As such, un-

derstanding and deﬁning the bounds of autonomy and

alignment becomes essential for effective system op-

eration. For this purpose, we adopt a pragmatic per-

spective on both autonomy and alignment.

Table 1: Matrix showcasing the interplay between grada-

tions of alignment (vertical) and autonomy (horizontal) in

the context of LLM-powered multi-agent architectures.

Levels of

Autonomy &

Alignment

L0: Static L1: Adaptive L2: Self-

Organizing

L2: Real-time

Responsive

User-

Supervised

Automation

User-

Collaborative

Adaptation

User-

Responsive

Autonomy

L1: User-

Guided

User-Guided

Automation

User-Guided

Adaptation

User-Guided

Autonomy

L0: Integrated Rule-Driven

Automation

Pre-

Conﬁgured

Adaptation

Bounded

Autonomy

3.1.1 Autonomy

The degree of autonomy refers to the extent to which

an AI system can make decisions and act indepen-

dently of rules and mechanisms deﬁned by humans.

For LLM-powered multi-agent systems, this trans-

lates to a system’s proﬁciency in addressing the goals

or tasks speciﬁed by the user in a self-organizing man-

ner, adapting and re-calibrating to the complexities

of a given situation. Autonomous multi-agent sys-

tems are by nature striving for this end-to-end auto-

matic goal completion and task management from a

user perspective. Automation pertains to tasks being

carried out without human input (Brustoloni, 1991;

SAE International, 2016), while autonomy pertains

to decisions about tasks being made without human

intervention (Franklin and Graesser, 1996; Parasura-

man et al., 2000; Beer et al., 2014). In the domain of

LLM-powered multi-agent systems, we look beyond

mere task automation, focusing on how these systems

internally manage their dynamics to fulﬁll user objec-

tives. Our taxonomy, therefore, distinguishes systems

on a spectrum of autonomy. Drawing from the triadic

interplay (Fig. 3), on the one end of the spectrum, we

see systems that heavily rely on predeﬁned rules, set

by system architects. While they may execute tasks

autonomously, their decision-making process is con-

strained within a ﬁxed set of parameters (low auton-

omy). On the other hand, we encounter systems char-

acterized by their ability for self-organisation and dy-

namic self-adaptation. Rather than relying on hard-

coded mechanisms, they harness the power of LLMs

to interpret, decide, and act, making them more adapt-

able to changing situations (high autonomy).

Autonomy Levels. The levels of autonomy, rep-

resented on the x-axis in our matrix (see Fig. 1

and Table 1), articulate the degree of agency of the

LLM-powered agents in making decisions regarding

the system operation, independently from predeﬁned

mechanisms.

L0: Static Autonomy - At this foundational level,

systems are primarily automated, relying heav-

ily on the rules, conditions, and mechanisms

embedded by system architects. The systems

follow deﬁned rules and predetermined mecha-

nisms. While the agents are not empowered to

modify these rules, some degree of ﬂexibility re-

mains resulting from rule-based options and al-

ternatives.

L1: Adaptive Autonomy - Evolving from the static

level, systems at this stage possess the capabil-

ity to adapt their behavior within a structure and

procedural guidelines established by the system

architects. The LLM-powered agents are capa-

ble of adjusting the system’s operations within

this provided framework (such as ﬂexible infras-

tructures and protocols) due to the needs of the

given application scenarios, but not beyond.

L2: Self-Organizing Autonomy - At this high-

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

est level of autonomy, LLM-powered agents

emerge as the principal actors, capable of self-

organization, actively learning and dynamically

tailoring their operations in real-time based on

environmental cues and experiences. However,

this might also include highly generic infrastruc-

tures that are modiﬁable by the LLM-powered

agents and thus allow self organisation.

3.1.2 Alignment

In the context of AI, the term alignment traditionally

refers to the challenge of ensuring that an AI sys-

tem’s behavior aligns with human intentions, values

or goals. This intricate problem, often framed as the

control problem, is a cornerstone of AI safety dis-

course (Bostrom, 2017; Russell, 2019). However,

when viewed through a practical lens, especially in

the context of autonomous LLM-powered multi-agent

systems, the alignment paradigm acquires a more

interactive, user-centric perspective (Amodei et al.,

2016), as it can be seen as a calibration of condi-

tions tied to user-prompted goals. This includes pref-

erences, policies, constraints, and boundaries which

collectively steer or regulate the system’s trajectory

towards achieving its set targets. Importantly, within

this framework, alignment is not seen as counter to

autonomy. Instead, it acts to complement and reﬁne

it, being applicable across various levels of autonomy.

For our taxonomy, we combine two important di-

mensions of alignment: its origin and timing, reﬂect-

ing the dynamic tension between automated align-

ment mechanisms and human customization, as illus-

trated in Fig. 3. The origin delves into who dictates

the alignment, the system architect or the system user.

Meanwhile, timing refers to when the alignment is

speciﬁed, encompassing phases like pre-deployment,

post-deployment but prior to runtime, or even during

runtime. Furthermore, we’ve categorized alignment

into levels. The base level, or low alignment level, sig-

niﬁes alignment that’s already embedded into the sys-

tem’s design by the system architects. This intrinsic

alignment sets broad behavioral boundaries without

focusing on speciﬁc user preferences. On the other

hand, the high alignment levels are more adaptable

and centered around user-speciﬁed alignment. Here,

users have the ﬂexibility to set their preferences either

before the system enters its runtime or, ultimately,

during its active operation.

Alignment Levels. The levels of alignment, repre-

sented on the y-axis in our matrix (see Fig. 1 and Ta-

ble 1), measure the degree to which users of the sys-

tem can inﬂuence or adjust the system’s behavior.

L0: Integrated Alignment - At this foundational

level, the alignment techniques are built directly

into the system’s architecture. In such system,

alignment mechanisms are static and rule-driven,

and cannot be altered by the users.

L1: User-Guided Alignment - Evolving from the

previous level, the User-Guided Alignment of-

fers a degree of customization. This level em-

powers users by allowing them to set or ad-

just speciﬁc alignment parameters, such as con-

ditions, rules, or boundaries, before the system

starts its operation. These interactions are pri-

marily facilitated via user interfaces designed to

capture user preferences in a structured manner.

L2: Real-Time Responsive Alignment - The high-

est level of alignment is represented by means to

adjust the system’s behavior in real-time. Thanks

to integrated real-time monitoring mechanisms,

the system can actively solicit user feedback user

decisions at critical junctures or decision points.

This responsiveness enables a high level of col-

laboration in terms of ongoing feedback between

the user and the system.

3.1.3 Combinations of Autonomy and Alignment

By combining these two dimensions in our matrix,

we provide a comprehensive view of the interplay be-

tween diverse gradations of autonomy and alignment

within LLM-powered multi-agent systems. Table 1

gives an overview of the employed levels and the re-

sulting spectrum of potential combinations.

Alignment

rule-driven

automation

user-guided

adaptation

user-responsive

autonomy

Automation

Autonomy

Balance

L1 L2

Figure 4: Interplay between autonomy and alignment: bal-

ancing evolving levels of dynamism and responsibilities of

both LLM-powered agents (autonomy) and human users

(alignment).

As illustrated in Fig. 4, departing from static and

rule-driven system conﬁgurations (automation), this

autonomy-alignment matrix captures the progression

of dynamism and responsibilities as we move along

the axes. On the y-axis, alignment levels represent

the gradation of human users’ involvement—from

integrated systems where the user’s role is passive

(L0), to real-time responsive setups demanding ac-

tive participation (L2). On the y-axis, the auton-

omy levels signify the evolving capabilities of LLM-

powered agents, progressing from static behaviors

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems

(L0) to adaptive (L1) and, ultimately, self-organizing

mechanisms (L2). This matrix structure reﬂects the

triadic interplay and dynamic tensions illustrated in

Fig. 3. As we delve deeper into the matrix, the chal-

lenge becomes evident: ensuring balance between the

evolving responsibilities of LLM-powered agents and

the goals and intentions by the human users, ulti-

mately resulting in a dynamic collaboration between

agents and humans.

3.2 Architectural Viewpoints

Architectural viewpoints are a structured means to an-

alyze and assess complex systems from diverse per-

spectives focusing on selected aspects and layers of

an architecture (Bass et al., 2003; Clements et al.,

2003). Central to these viewpoints is the considera-

tion of stakeholder concerns, which inform and de-

termine the highlighted aspects and their interrela-

tions in each viewpoint. Providing a combined multi-

perspective analysis, viewpoints serve as an effective

framework to examine the structures and dynamics of

software architectures. For our taxonomy, we lever-

age viewpoints on autonomous LLM-powered multi-

agent systems. Rather than mapping the autonomy-

alignment taxonomy ﬂatly onto the system, which

oversimpliﬁes the multi-faceted nature of these sys-

tems, analyzing each architectural viewpoint individ-

ually offers a tailored lens, enabling to comprehend

the role and impact of autonomy and alignment within

the system. Each viewpoint reveals distinct insights

into the system’s behavior, internal interactions, com-

position, and context interaction, leading to a more

nuanced and comprehensive classiﬁcation (Rozanski

and Woods, 2012).

«uses»

Functional

Viewpoint

Goal-driven

Task Mgmt.

Development

Viewpoint

Agent

Composition

Multi-Agent

Collaboration

Process

Viewpoint

Context

Interaction

Physical

Viewpoint

Figure 5: Architectural viewpoints oriented to the 4+1 view

model of software architecture (Kruchten, 1995) applied to

autonomous LLM-powered multi-agent systems.

3.2.1 Applied Viewpoints

For our taxonomy, we orient to Kruchten’s renowned

4+1 view model of software architecture (Kruchten,

1995), an established standard viewpoint model for

software architecture, adapting it to suit the archi-

tectural characteristics of LLM-powered multi-agent

systems (see Section 2.2). Our taxonomy encom-

passes the following four architectural viewpoints on

these systems (refer to Fig. 5 and 2):

G Goal-Driven Task Management (Functional

Viewpoint): Kruchten’s functional viewpoint

refers to the system’s visible functionalities

as experienced by its users (Kruchten, 1995).

In the context of autonomous LLM-powered

multi-agent systems, we see Goal-driven Task

Management as a manifestation of this functional

viewpoint. It entails the system’s capabilities and

mechanisms to decompose user-prompted goals

or complex tasks into manageable tasks, and sub-

sequently, orchestrate task execution, combine the

results, and deliver the ﬁnal result forming the re-

sponse.

A Agent Composition (Development Viewpoint):

According to Kruchten, the development view-

point is primarily focusing on the system’s soft-

ware architecture, the breakdown into compo-

nents, and their organization (Kruchten, 1995).

In our context, we interpret this as Agent

Composition, focusing on the system’s internal

composition, particularly the assembly and con-

stellation of agents. It includes the types and roles

of agents, their memory usage, the relationships

between agents.

M Multi-Agent Collaboration (Process Viewpoint):

Kruchten’s process viewpoint concerns the dy-

namic aspects of a system, speciﬁcally the sys-

tem procedures and interactions between com-

ponents (Kruchten, 1995). We apply this to

the Multi-Agent Collaboration in our model,

emphasizing the collaborative task execution and

interactions among agents. This encompasses the

application of communication protocols, the dy-

namics of actions management, such as the actual

task execution, mutual task delegation, as well as

the evaluation and merging of task results on agent

level, as well as the management of communica-

tion components and prompt engineering.

C Context Interaction (Physical Viewpoint): Ac-

cording to Kruchten, the physical viewpoint in-

volves the system’s mapping to physical resources

(Kruchten, 1995). We extend this to Context

Interaction, focusing on the system’s interac-

tion with the external environment. It includes

how the system acquires, integrates, and utilizes

contextual resources such as external data, expert

tools, and further foundation models as well as

the organized distribution and utilization of con-

textual resources within the agent network.

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

alternatives

mandatory

«adapts to»

Goal-driven

Task Mgmt.

Multi-Agent

Collaboration

Agent

Composition

Context

Interaction

Decomposition

Goal-driven

Task Mgmt.

Orchestration Synthesis

Agent

Composition

Agent

Generation

Role

Definition

Memory

Usage

Network

Management

Multi-Agent

Collaboration

Communication

Protocol Mgmt.

Prompt

Engineering

Action

Management

Autonomy

Level

Static

(L0)

Adaptive

(L1)

Self-Organizing

(L2)

Alignment

Level

Integrated

(L0)

User-Guided

(L1)

Real-Time

Respons. (L2)

(a)

(b) (c)

(d)

(e)

Multi-Agent

Architectural Viewpoint

Context

Interaction

Resources

Integration

Resources

Utilization

Figure 6: Feature diagram showcasing the taxonomic structure. Each viewpoint integrates autonomy and alignment levels (a).

The diagram further illustrates viewpoint-speciﬁc aspects and mechanisms (b–e).

3.3 Interplay of Autonomy and

Alignment in the System

Architecture

As already illustrated, both autonomy and alignment

serve as cross-cutting concerns (Kiczales et al., 1997)

impacting the operational efﬁciency of various ar-

chitectural aspects across LLM-powered multi-agent

systems. Thus, in the following, we map our ma-

trix of autonomy and alignment levels onto the ar-

chitectural viewpoints. This projection crafts a three-

dimensional matrix, offering a prism through which

these systems can be analyzed and categorized (also

see Fig. 1). (Händler, 2023) provides a systematic

overview of the resulting viewpoint-speciﬁc combi-

nations of autonomy and alignment levels.

Viewpoint-Speciﬁc Aspects and Level Criteria.

As outlined above, architectural viewpoints provide

means to analyze certain aspects and aspect relations

of the system’s architecture in a multi-perspective

manner (Rozanski and Woods, 2012). We now sys-

tematize the viewpoint-speciﬁc aspects employed in

our taxonomy.

Fig. 6 gives an overview of our taxonomy’s char-

acteristics, structured through a feature diagram (Ba-

tory, 2005; Schobbens et al., 2007). In particu-

lar, Fig. 6 (a) structures the viewpoint-speciﬁc taxo-

nomic structure. Each of the four integrated view-

points provides a certain combination of autonomy

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems

and alignment levels. As illustrated in Figs. 6 (b–e),

this structure is reﬁned by viewpoint-speciﬁc aspects

and their interdependencies in terms of requirements-

driven dependencies (adapts-to), presuming a high-

autonomy system conﬁguration (Händler, 2023).

These dependencies suggest that the capabilities of a

dependent aspect evolve in line with the needs and

stipulations of the aspect it points to. In turn, also

these viewpoint-speciﬁc aspects can be assessed by

the autonomy and alignment levels, resulting in a

more nuanced taxonomic classiﬁcation.

For a comprehensive speciﬁcation of level crite-

ria for both autonomy and alignment related to each

aspect, as well as for a discussion of the interdepen-

dencies among these aspects, please refer to (Händler,

2023).

4 CLASSIFICATION OF

SELECTED SYSTEMS

In order to demonstrate the practical utility of our

taxonomy, we analyze and classify selected existing

autonomous LLM-powered multi-agent systems. We

have chosen a set of seven state-of-the-art multi-agent

systems for this assessment: AUTOGPT (Torantulino

et al., 2023), BABYAGI (Nakajima, 2023), SUPER-

AGI (TransformerOptimus et al., 2023), HUGGING-

GPT (Shen et al., 2023), METAGPT (Hong et al.,

2023), CAMEL (Li et al., 2023), and AGENTGPT

(Shrestha et al., 2023). Each of these systems is main-

tained and available as open-source project. For each

selected system, we gathered relevant information by

examining the technical documentation and research

papers, where available, as well as reviewing the code

base. We further engaged with each system to explore

its real-time functionalities, with emphasis on align-

ment mechanisms available before and during run-

time.

Taxonomic Classiﬁcation. The taxonomic classiﬁ-

cation relies on a detailed assessment of autonomy

and alignment levels for viewpoint-speciﬁc aspects of

the systems. Table 2 reports on the results of assess-

ing these levels of autonomy (AU) and alignment (AL)

for aspects characterizing the four architectural view-

points applied by our taxonomy. In particular, for

Goal-driven Task Management, the aspects of de-

composition (Decom), orchestration (Orch), and syn-

thesis (Synth); for Multi-Agent Collaboration,

the aspects of communication-protocol management

(CommP), prompt engineering (PrEng), and action

management (ActM); for Agent Composition, the

aspects of agent generation (AGen), role deﬁnition

(RoleD), memory usage (MemU), and network manage-

ment (NetM); for Context Interaction, the aspects

of resource integration (Integ), and resource utiliza-

tion Util are distinguished. The viewpoint-speciﬁc

aspects with corresponding level criteria applied for

the assessment are detailed in (Händler, 2023).

Fig. 7 displays the derived autonomy and align-

ment levels per multi-agent system using radar (or

spider) charts (Tufte, 2001). In particular, archi-

tectural aspects form the multiple axes. The level

scheme (L0, L1, L2) for autonomy and alignment is

depicted by grey circles linking these axes. The blue

graph then represents the assessed autonomy levels,

the green dashed graph the corresponding alignment

levels. The extended paper version (Händler, 2023)

provides a detailed discussion of analysis results per

system as well as a comprehensive comparative anal-

ysis.

Strategies Across System Groups. We now explore

how different categories of systems balance the inter-

play between autonomy and alignment. Based on our

taxonomic classiﬁcation and the resulting system pro-

ﬁles as illustrated in Fig. 7, we can categorize the se-

lected 7 systems under analysis into three distinct sys-

tem groups, which encompass general-purpose sys-

tems, central-controller systems, and role-agent sys-

tems. It’s important to note that our categorization

into these three groups, based on the systems chosen

for this exploration, doesn’t capture the entire spec-

trum of autonomous LLM-powered multi-agent sys-

tems. For a comprehensive overview of existing sys-

tems, we recommend referring to the recent surveys

provided by (Wang et al., 2023; Xi et al., 2023). In the

following, the key characteristics as observed from

the corresponding system proﬁles are discussed.

• General-Purpose Systems - representing multi-

agent systems designed for and adaptable to a

broad spectrum of tasks and applications. Within

the analyzed set of multi-agent systems, the fol-

lowing fall into this group: AUTO-GPT (Toran-

tulino et al., 2023), BABYAGI (Nakajima, 2023),

SUPERAGI (TransformerOptimus et al., 2023),

and AGENTGPT (Shrestha et al., 2023). Goals

are decomposed autonomously and represented

as prioritized task lists (L2 Decom). They em-

ploy a multi-cycle process framework performed

by dedicated task-management agents represented

by certain generic agent types, including a single

task-execution agent. Relations and communica-

tions between these agents are strictly predeﬁned,

and agent conversations express as a monologue

of the task-execution agent, resulting in low au-

tonomy levels (L0) for communication protocol

(CommP), and network management (NetM). The

task-related actions are performed autonomously

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

Table 2: Assessment of autonomy (AU) and alignment (AL) levels across viewpoint-speciﬁc aspects of selected LLM-powered

multi-agent systems. Detailed level criteria for viewpoint-speciﬁc aspects are discussed in (Händler, 2023). * ZAPIER, a

workﬂow-automation tool, has been included to contrast the results.

LLM-powered

Multi-Agent

Systems

Goal-driven Task Mgmt. Multi-Agent Collaboration Agent Composition Context Interact.

Decom Orch Synth CommP PrEng ActM AGen RoleD MemU NetM Integ Util

AU AL AU AL AU AL AU AL AU AL AU AL AU AL AU AL AU AL AU AL AU AL AU AL

Auto-GPT (Torantulino et al., 2023) 2 0 0 0 1 0 0 0 1 0 2 0 0 0 1 0 0 0 0 0 0 0 2 0

BabyAGI (Nakajima, 2023) 2 0 0 0 1 0 0 0 1 0 2 0 0 0 1 0 0 0 0 0 0 0 2 0

SuperAGI (TransformerOptimus et al., 2023) 2 0 1 0 1 1 0 0 1 0 2 0 1 1 2 1 0 1 0 0 0 1 2 1

HuggingGPT (Shen et al., 2023) 2 0 1 0 2 0 0 0 2 0 2 0 2 0 2 0 1 0 0 0 2 0 2 0

MetaGPT (Hong et al., 2023) 2 0 0 0 2 0 1 0 1 0 2 0 0 0 0 0 0 0 1 0 0 0 2 0

CAMEL (Li et al., 2023) 2 0 0 0 1 0 0 0 1 0 1 0 0 1 1 1 0 0 0 1 0 0 0 0

AgentGPT (Shrestha et al., 2023) 2 1 1 0 1 0 0 0 1 0 2 0 1 1 2 0 0 0 0 0 0 0 2 1

Zapier* (Rahmati et al., 2017) 1 1 0 1 0 1 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 1 0 1

by the task-execution agent (mostly L2 auton-

omy ActM). While resource integration is based

on provided mechanisms (Integ), the resources

are selected and utilized by the LLM-powered

in a self-organizing manner (L2 autonomy for

Util), except for CAMEL; resulting in similar

autonomy proﬁles for the aforementioned aspects.

Besides from these commonalities, these sys-

tems distinguish in certain characteristics. Both

AUTO-GPT and BABYAGI employ generic task-

execution agent, and provide no further align-

ment options at all. Moreover, these systems em-

ploy a generic task-execution agent with prede-

ﬁned agent roles and relations, resulting in L0 au-

tonomy for AGen and NetM. In contrast, SUPER-

AGI and AGENTGPT employ execution agents

with self-organizing agent roles (L2 autonomy for

RoleD), an adaptable orchestration process (L1

for Orch), and some alignment options, especially

for agent-speciﬁc aspects. Moreover, these sys-

tems employ execution agents, whose roles can be

customized by the user (L1 alignment for AGen).

• Central LLM Controller - marks a third group

specialized in leveraging and combining contex-

tual resources for accomplishing complex goals.

HUGGINGGPT (Shen et al., 2023) serves as an

archetype of such systems, utilizing resources es-

pecially in terms of existing ML models inte-

grated via HUGGING FACE. HUGGINGGPT is

characterized by a single central LLM-powered

control agent with monologue-based reﬂection

and planning. Language serves as generic in-

terface to manage the interplay between multi-

ple specialized foundation models. In compari-

son to other systems or system groups, we see the

highest levels of autonomy granted to this central

agent (mostly L2); also see Fig. 7 (d). Further-

more, we see a ﬁnite and artefact-oriented process

adaptable by the LLM-powered agent for orches-

trating the different model-related tasks (L1 au-

tonomy). Beyond prompting the task, there are no

further user-centric alignment options (L0 align-

ment).

• Role-Agent Systems - employ an interplay

or simulation between multiple dedicated roles

agents. This collaboration can serve different pur-

poses, such as simulating a discussion or solving

tasks that demand for a multi-perspective collabo-

ration. With deﬁned roles in a certain environment

(such as in a software development project), their

application is bound to this application domain or

special purpose. Among the analyzed systems,

METAGPT (Hong et al., 2023) and CAMEL (Li

et al., 2023) represent such systems. In con-

trast to the general-purpose systems, the execution

agents play roles with dedicated responsibilities in

a certain application domain. Furthermore, these

role agents actually collaborate directly with each

other. In case of the two exemplary systems, this

collaboration is realized by communication pro-

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

SuperAGI

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

CAMEL

(f)

(c)

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

BabyAGI

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

Zapier*

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

MetaGPT

(h)

(e)

(b)

Decom

level scheme (L0, L1, L2)

autonomy level

alignment level

architectural aspect

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

Auto-GPT

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

HuggingGPT

Decom

Orch

Synth

CommP

PrEng

ActM

AGen

RoleD

MemU

NetM

Integ

U�l

AgentGPT

(a)

(d)

(g)

Figure 7: Radar charts illustrating the system proﬁles based on an assessment of architectural aspects in terms of autonomy

(blue graph) and alignment (green dashed graph) levels. Detailed assessment data can be found in Table 2.

tocols employing a dynamic exchange between

agents with instructor and executor roles. In par-

ticular, CAMEL employs two such role agents

based on predeﬁned agent types, but adjustable

by the user. In ongoing strict dialogue cycles,

the AI-user role agents instructs the AI-assistant

role agent to execute the tasks (L0 autonomy for

CommP). Similar to SUPERAGI, CAMEL requires

the user to specify the agents’ roles (L1 align-

ment). METAGPT, in contrast, internally assigns

predeﬁned roles with responsibilities alongside

a waterfall development process (L0 alignment);

thus, also expressing a ﬁnite and artefact-oriented

process (L0 autonomy for Orch), terminating with

the produced and tested software program. How-

ever, like in real-world software project, reﬁne-

ment iterations can follow, optional feedback cy-

cles make it adaptable for the agents (L1 auton-

omy for CommP).

Strategy Assessment. Beyond differences in the ap-

plied communication protocols, it is the ﬂexibility of

agent roles (in relation to both autonomy and align-

ment) and further customization options for agent-

speciﬁc aspects that distinguishes the systems’ strate-

gies (see above). However, when examining how the

systems deal with autonomy and alignment across

further aspects, most systems and system groups

show similar strategies. The reasoning capabilities of

LLM-powered agents are especially leveraged in ar-

eas demanding high autonomy, such as the goal de-

composition, the actual execution of task-related ac-

tions, and the utilization of contextual resources. In-

terestingly, these high-autonomy aspects are mostly

combined with low alignment levels, resulting in

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

bounded autonomy aspects (refer to Table 1). A

closer look at aspect interdependencies, as depicted

in Fig. 6, reveals that these internally unbalanced as-

pects are accompanied by other low-autonomy as-

pects equipped with limited ﬂexibility. For instance,

autonomous action management depends on strict or

predeﬁned communication protocol. For further de-

tails, refer to (Händler, 2023). In these cases, the

predeﬁned and rule-based mechanisms serve as inte-

grated alignment guiding and controlling the accurate

operation of the dependent autonomous aspects.

5 CONCLUSION

In this paper, we have introduced a comprehensive

multi-dimensional taxonomy engineered to analyze

how autonomous LLM-powered multi-agent systems

balance the dynamic interplay between autonomy and

alignment across their system architectures. For this

purpose, the taxonomy employs a matrix that com-

bines hierarchical levels of autonomy and alignment.

This matrix is then mapped onto various architec-

tural aspects organized by four architectural view-

points reﬂecting different complementary concerns

and perspectives. The resulting taxonomic system en-

ables the assessment of interdependent aspect con-

ﬁgurations in a wide spectrum, ranging from simple

conﬁgurations, such as predeﬁned mechanisms com-

bined with system-integrated alignment (rule-driven

automation), to sophisticated conﬁgurations, such as

self-organizing agency responsive to user feedback

and evolving conditions (user-responsive autonomy).

Applied to 12 distinct architectural aspects inher-

ent to viewpoints, such as goal-driven task man-

agement, multi-agent collaboration, agent composi-

tion, and context interaction, this taxonomy allows

for a nuanced analysis and understanding of architec-

tural complexities within autonomous LLM-powered

multi-agent systems.

Through our taxonomy’s application to seven se-

lected LLM-powered multi-agent systems, its practi-

cal relevance and utility has been illustrated. In par-

ticular, it has been shown that a combined assess-

ment of autonomy and alignment levels across the ar-

chitectural aspects of each multi-agent system allows

for identifying system proﬁles that can indicate cer-

tain strategies for balancing the dynamic interplay be-

tween autonomy and alignment. This exploration of

exemplary current systems also revealed several chal-

lenges, which are detailed in (Händler, 2023).

For additional analyses, a more comprehensive discus-

sion and extended results, readers are referred to the ex-

tended paper version available on arXiv (Händler, 2023).

Most prominently, we observed a lack of user-

centric alignment options across all systems, with lit-

tle user-guided alignment, but no real-time responsive

alignment at all. Moreover, the systems exhibit high

autonomy levels mostly for certain aspects, such as

the goal decomposition, the action management, or

the utilization of contextual resources. In contrast,

other key aspects of the system operation show lim-

ited autonomy; aspects such as managing the commu-

nication protocol, memory usage, or agent network

are largely static, leaning heavily on predeﬁned mech-

anisms.

Based on these and further ﬁndings, we espe-

cially see two promising avenues for the evolu-

tion of autonomous LLM-powered multi-agent sys-

tems. Firstly, by employing adaptable and self-

organizing communication protocols and agent net-

works, the systems’ role-playing capabilities could be

enhances, which enables them to better simulate com-

plex multi-perspective environments. By reﬂecting

diverse standpoints and strategies, this could also pave

the way for more in-depth inter-agent discussions and

creativity in problem solving. Secondly, the explo-

ration of real-time responsive systems, which can

adapt to evolving conditions as well as to user feed-

back during runtime, would foster dynamic collabo-

ration and hybrid teamwork between LLM-powered

agents and human users.

Departing from an exploratory stage, the ﬁeld

of autonomous LLM-powered multi-agent systems is

rapidly evolving, resulting in a growing number of

promising approaches and innovative architectures.

With their current capabilities and inherent potentials,

such as multi-perspective domain simulations or col-

laborative environments of autonomous agents and

human coworkers, these systems could signiﬁcantly

contribute to the progression towards advanced stages

of artiﬁcial intelligence, such as AGI or ASI. From a

pragmatic perspective, there are numerous opportuni-

ties for combining LLMs as general purpose technol-

ogy with the speciﬁcs of various application domains.

LLM-based multi-agent systems can serve as foun-

dation for developing corresponding domain-speciﬁc

application layers. The architectural complexities re-

sulting from the dynamic interplay between auton-

omy and alignment can be seen as one of the key

challenges in such systems. By providing a system-

atic framework for analyzing these complexities, our

taxonomy aims to contribute to these ongoing efforts.

For our subsequent endeavors, we aim at devel-

oping a comprehensive overview and comparison of

existing autonomous LLM-powered multi-agent sys-

tems, complementing existing literature reviews in the

ﬁeld (Wang et al., 2023; Xi et al., 2023). To this end,

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems

we intend to analyze and classify available systems

using our taxonomy. The identiﬁed system proﬁles

and balancing strategies resulting from this analysis

will then be combined with further investigations of

functional system capabilities.

Building on the foundation of our taxonomy, fu-

ture initiatives could venture into the following areas:

A dedicated exploration and systematization of align-

ment techniques, particularly tailored for LLM-based

interaction and application layers, could serve as ref-

erence for future systems. Moreover, the conception

of a methodological framework with instruments and

benchmarks for measuring the functional capabilities

of LLM-powered multi-agent systems could provide

a structured template to evaluate key metrics like efﬁ-

ciency, accuracy, and scalability of these systems.

ACKNOWLEDGEMENTS

The author gratefully acknowledges the support from

the "Gesellschaft für Forschungsförderung (GFF)" of

Lower Austria, as this research was conducted at Fer-

dinand Porsche Mobile University of Applied Sci-

ences (FERNFH) as part of the "Digital Transforma-

tion Hub" project funded by the GFF.

REFERENCES

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schul-

man, J., and Mané, D. (2016). Concrete problems in

AI safety. arXiv preprint arXiv:1606.06565.

Bass, L., Clements, P., and Kazman, R. (2003). Software ar-

chitecture in practice. Addison-Wesley Professional.

Batory, D. (2005). Feature models, grammars, and proposi-

tional formulas. In 9th International Software Product

Line Conference, pages 7–20.

Beer, J. M., Fisk, A. D., and Rogers, W. A. (2014). Toward

a framework for levels of robot autonomy in human-

robot interaction. Journal of human-robot interaction,

3(2):74.

Bird, S. D. (1993). Toward a taxonomy of multi-agent sys-

tems. International Journal of Man-Machine Studies,

39(4):689–704.

Bommasani, R., Hudson, D. A., Adeli, E., Altman, R.,

Arora, S., von Arx, S., Bernstein, M. S., Bohg, J.,

Bosselut, A., Brunskill, E., et al. (2021). On the

opportunities and risks of foundation models. arXiv

preprint arXiv:2108.07258.

Bostrom, N. (2017). Superintelligence. Dunod.

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., et al. (2020). Language models are few-

shot learners. Advances in neural information pro-

cessing systems, 33:1877–1901.

Brustoloni, J. C. (1991). Autonomous agents: Characteri-

zation and requirements. Carnegie Mellon University.

Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J.,

Horvitz, E., Kamar, E., Lee, P., Lee, Y. T., Li, Y.,

Lundberg, S., et al. (2023). Sparks of artiﬁcial general

intelligence: Early experiments with GPT-4. arXiv

preprint arXiv:2303.12712.

Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra,

G., Roberts, A., Barham, P., Chung, H. W., Sut-

ton, C., Gehrmann, S., et al. (2022). Palm: Scal-

ing language modeling with pathways. arXiv preprint

arXiv:2204.02311.

Clements, P., Garlan, D., Little, R., Nord, R., and Stafford,

J. (2003). Documenting software architectures: views

and beyond. In 25th International Conference on Soft-

ware Engineering, 2003. Proceedings., pages 740–

741. IEEE.

Dudek, G., Jenkin, M. R., Milios, E., and Wilkes, D. (1996).

A taxonomy for multi-agent robotics. Autonomous

Robots, 3:375–397.

Fabiano, F., Pallagani, V., Ganapini, M. B., Horesh, L.,

Loreggia, A., Murugesan, K., Rossi, F., and Srivas-

tava, B. (2023). Fast and slow planning. arXiv preprint

arXiv:2303.04283.

Franklin, S. and Graesser, A. (1996). Is it an agent, or just a

program?: A taxonomy for autonomous agents. In In-

ternational workshop on agent theories, architectures,

and languages, pages 21–35. Springer.

Hong, S., Zheng, X., Chen, J., Cheng, Y., Zhang, C.,

Wang, Z., Yau, S. K. S., Lin, Z., Zhou, L., Ran,

C., et al. (2023). MetaGPT: Meta programming for

multi-agent collaborative framework. arXiv preprint

arXiv:2308.00352.

Händler, T. (2023). Balancing autonomy and alignment: A

multi-dimensional taxonomy for autonomous LLM-

powered multi-agent architectures. arXiv preprint

arxiv:2310.03659. https://doi.org/10.48550/arXiv.

2310.03659.

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E.,

Bang, Y. J., Madotto, A., and Fung, P. (2023). Survey

of hallucination in natural language generation. ACM

Computing Surveys, 55(12):1–38.

Kaddour, J., Harris, J., Mozes, M., Bradley, H., Raileanu,

R., and McHardy, R. (2023). Challenges and ap-

plications of large language models. arXiv preprint

arXiv:2307.10169.

Kahneman, D. (2011). Thinking, fast and slow. Macmillan.

Kiczales, G., Lamping, J., Mendhekar, A., Maeda,

C., Lopes, C., Loingtier, J.-M., and Irwin, J.

(1997). Aspect-oriented programming. In

ECOOP’97—Object-Oriented Programming: 11th

European Conference Jyväskylä, Finland, June 9–13,

1997 Proceedings 11, pages 220–242. Springer.

Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., and Iwasawa, Y.

(2022). Large language models are zero-shot reason-

ers. Advances in neural information processing sys-

tems, 35:22199–22213.

Kruchten, P. B. (1995). Architectural blueprints — the

“4+1” view model of software architecture. IEEE soft-

ware, 12(6):42–50.

A Taxonomy for Autonomous LLM-Powered Multi-Agent Architectures

Li, G., Hammoud, H. A. A. K., Itani, H., Khizbullin, D., and

Ghanem, B. (2023). CAMEL: Communicative agents

for "mind" exploration of large scale language model

society. arXiv preprint arXiv:2303.17760.

Lin, B. Y., Fu, Y., Yang, K., Ammanabrolu, P., Brahman,

F., Huang, S., Bhagavatula, C., Choi, Y., and Ren,

X. (2023). SwiftSage: A generative agent with fast

and slow thinking for complex interactive tasks. arXiv

preprint arXiv:2305.17390.

Maes, P. (1995). Artiﬁcial life meets entertainment: life-

like autonomous agents. Communications of the ACM,

38(11):108–114.

Maynez, J., Narayan, S., Bohnet, B., and McDonald, R.

(2020). On faithfulness and factuality in abstractive

summarization. arXiv preprint arXiv:2005.00661.

Minsky, M. (1988). The Society of mind. Simon and Schus-

ter.

Mintzberg, H. (1989). The structuring of organizations.

Springer.

Moya, L. J. and Tolk, A. (2007). Towards a taxonomy of

agents and multi-agent systems. In SpringSim (2),

pages 11–18.

Nakajima, Y. (2023). BabyAGI. https://github.com/

yoheinakajima/babyagi.

Narendra, K. S. and Annaswamy, A. M. (2012). Stable

adaptive systems. Courier Corporation.

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright,

C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K.,

Ray, A., et al. (2022). Training language models to

follow instructions with human feedback. Advances

in Neural Information Processing Systems, 35:27730–

27744.

O’reilly Iii, C. A. and Tushman, M. L. (2008). Ambidex-

terity as a dynamic capability: Resolving the innova-

tor’s dilemma. Research in organizational behavior,

28:185–206.

Parasuraman, R., Sheridan, T. B., and Wickens, C. D.

(2000). A model for types and levels of human in-

teraction with automation. IEEE Transactions on sys-

tems, man, and cybernetics-Part A: Systems and Hu-

mans, 30(3):286–297.

Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang,

P., and Bernstein, M. S. (2023). Generative agents: In-

teractive simulacra of human behavior. arXiv preprint

arXiv:2304.03442.

Rahmati, A., Fernandes, E., Jung, J., and Prakash, A.

(2017). IFTTT vs. Zapier: A comparative study

of trigger-action programming frameworks. arXiv

preprint arXiv:1709.02788.

Rozanski, N. and Woods, E. (2012). Software systems

architecture: working with stakeholders using view-

points and perspectives. Addison-Wesley.

Russell, S. (2019). Human compatible: Artiﬁcial intelli-

gence and the problem of control. Penguin.

Russell, S. (2022). Artiﬁcial intelligence and the problem of

control. Perspectives on Digital Humanism, page 19.

Russell, S., Dewey, D., and Tegmark, M. (2015). Re-

search priorities for robust and beneﬁcial artiﬁcial in-

telligence. AI magazine, 36(4):105–114.

SAE International (2016). Taxonomy and deﬁnitions for

terms related to driving automation systems for on-

road motor vehicles.

Schobbens, P.-Y., Heymans, P., Trigaux, J.-C., and Bon-

temps, Y. (2007). Generic semantics of feature dia-

grams. Computer networks, 51(2):456–479.

Shen, Y., Song, K., Tan, X., Li, D., Lu, W., and Zhuang,

Y. (2023). HuggingGPT: Solving AI tasks with Chat-

GPT and its friends in Hugging Face. arXiv preprint

arXiv:2303.17580.

Shrestha, A., Subedi, S., and Watkins, A. (2023). Agent-

GPT. https://github.com/reworkd/AgentGPT.

Sloman, S. A. (1996). The empirical case for two systems

of reasoning. Psychological bulletin, 119(1):3.

Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kul-

shreshtha, A., Cheng, H.-T., Jin, A., Bos, T., Baker,

L., Du, Y., et al. (2022). Lamda: Language models for

dialog applications. arXiv preprint arXiv:2201.08239.

Torantulino et al. (2023). Auto-GPT. https://github.com/

Signiﬁcant-Gravitas/Auto-GPT.

Tosic, P. T. and Agha, G. A. (2004). Towards a hierarchical

taxonomy of autonomous agents. In 2004 IEEE In-

ternational Conference on Systems, Man and Cyber-

netics (IEEE Cat. No. 04CH37583), volume 4, pages

3421–3426. IEEE.

TransformerOptimus et al. (2023). SuperAGI. https://

github.com/TransformerOptimus/SuperAGI.

Tufte, E. R. (2001). The visual display of quantitative in-

formation, volume 2. Graphics press Cheshire, CT.

Van Dyke Parunak, H., Brueckner, S., Fleischer, M., and

Odell, J. (2004). A design taxonomy of multi-agent

interactions. In Agent-Oriented Software Engineer-

ing IV: 4th InternationalWorkshop, AOSE 2003, Mel-

bourne, Australia, July 15, 2003. Revised Papers 4,

pages 123–137. Springer.

Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang,

J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al. (2023).

A survey on large language model based autonomous

agents. arXiv preprint arXiv:2308.11432.

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F.,

Chi, E., Le, Q. V., Zhou, D., et al. (2022). Chain-of-

thought prompting elicits reasoning in large language

models. Advances in Neural Information Processing

Systems, 35:24824–24837.

Wolf, Y., Wies, N., Levine, Y., and Shashua, A. (2023).

Fundamental limitations of alignment in large lan-

guage models. arXiv preprint arXiv:2304.11082.

Wooldridge, M. and Jennings, N. R. (1995). Intelligent

agents: Theory and practice. The knowledge engineer-

ing review, 10(2):115–152.

Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B.,

Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2023).

The rise and potential of large language model based

agents: A survey. arXiv preprint arXiv:2309.07864.

Yudkowsky, E. (2016). The AI alignment problem: why it

is hard, and where to start. Symbolic Systems Distin-

guished Speaker, 4.

Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M.,

Chen, S., Dewan, C., Diab, M., Li, X., Lin, X. V., et al.

(2022). OPT: Open pre-trained transformer language

models. arXiv preprint arXiv:2205.01068.

KMIS 2023 - 15th International Conference on Knowledge Management and Information Systems