TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based

Innovation

Kamil Szczepanik

and Jarosław A. Chudziak

The Institute of Computer Science, Warsaw University of Technology, Poland

{kamil.szczepanik.stud, jaroslaw.chudziak}@pw.edu.pl

Keywords:

Large Language Model, LLM Agents, Multi-Agent Systems, TRIZ, Problem-Solving.

Abstract:

TRIZ, the Theory of Inventive Problem Solving, is a structured, knowledge-based framework for innovation

and abstracting problems to ﬁnd inventive solutions. However, its application is often limited by the complex-

ity and deep interdisciplinary knowledge required. Advancements in Large Language Models (LLMs) have

revealed new possibilities for automating parts of this process. While previous studies have explored single

LLMs in TRIZ applications, this paper introduces a multi-agent approach. We propose an LLM-based multi-

agent system, called TRIZ agents, each with specialized capabilities and tool access, collaboratively solving

inventive problems based on the TRIZ methodology. This multi-agent system leverages agents with various

domain expertise to efﬁciently navigate TRIZ steps. The aim is to model and simulate an inventive process

with language agents. We assess the effectiveness of this team of agents in addressing complex innovation

challenges based on a selected case study in engineering. We demonstrate the potential of agent collaboration

to produce diverse, inventive solutions. This research contributes to the future of AI-driven innovation, show-

casing the advantages of decentralized problem-solving in complex ideation tasks.

1 INTRODUCTION

Problem-solving is a fundamental subject of every in-

novation process. It is much more than ﬁnding quick

ﬁxes or solutions to occurring obstacles. Innovation

problem-solving requires substantial knowledge and

unconventional thinking. However, it is strongly ad-

vised to follow methodologies designed to succeed in

this process. Structured methodologies for problem-

solving in innovation, like TRIZ (Theory of Inven-

tive Problem Solving) (Orloff, 2006), Design Think-

ing (Brown et al., 2008), Lean Innovation (Sehested

and Sonnenberg, 2010), and more, were created.

Researchers have preferred or disliked methods;

some work only for speciﬁc problems or become hard

to implement. The innovation process is a challenging

task, and every tool that tries to harness it is merely a

rough guide for innovators during the process. Never-

theless, these methodologies are the best-known way

to cope with it.

TRIZ has been around for decades, proving its ef-

ﬁciency and usefulness as a problem-solving method-

ology. It was developed after analyzing around

https://orcid.org/0009-0009-6807-2426

https://orcid.org/0000-0003-4534-8652

200,000 patents (Loh et al., 2006) to create a struc-

tured framework for engineers and innovators to solve

innovation problems. Over the years, many improve-

ments have been proposed; however, the general idea

has stayed the same. In addition, complementary

tools and group work techniques have been devel-

oped to assist teams during the TRIZ process, such

as brainstorming (Mahto, 2013).

Although the TRIZ methodology has been suc-

cessfully helping innovators solve problems, it has

limitations that still need to be addressed. The biggest

one is perhaps the need for experts. The only way to

solve complex innovation challenges is for innovators

to have extensive knowledge and experience in the

relevant ﬁeld (Czinki and Hentschel, 2016) (Ilevbare

et al., 2013). It takes many years to become an ex-

pert in one domain, and sometimes, expertise in mul-

tiple domains is required to solve complex innovation

problems. People working in an innovation team must

think creatively and possess broad knowledge. There

is a need for interdisciplinary knowledge when cre-

ating a TRIZ team. There have been proposed con-

cepts supporting knowledge discovery (Choinski and

Chudziak, 2009); however, with the advent of LLMs,

new possibilities have arisen.

This study simulates a realistic team tasked with

196

Szczepanik, K. and Chudziak, J. A.

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation.

DOI: 10.5220/0013321900003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 196-207

ISBN: 978-989-758-737-5; ISSN: 2184-433X

solving innovation problems using the TRIZ method-

ology. It aims to model agents’ subtasks, goals, and

abilities for collaboratively achieving a theoretical so-

lution proposal. Choosing the TRIZ methodology

narrows the scope of actions while guiding the team

in a structured and systematic way.

The primary purpose is to showcase the abilities

of language agents in collaborative teamwork. This

study proposes a multi-agent system in which LLM

agents simulate a team’s workﬂow to solve innova-

tion problems using the TRIZ methodology. First, we

review previous works on this topic and introduce the

reader to the TRIZ methodology. Next, we describe

the workﬂow design, deﬁning key assumptions and

the system’s scope. The architecture and implemen-

tation details are discussed, outlining agent orchestra-

tion, agents’ abilities, and tools. The subject of the

system’s work is a case study of the TRIZ application

for the improvement design of a gantry crane (Luing

et al., 2024), which precisely documents each step of

the researchers’ workﬂow. Lastly, the system results

are assessed and discussed.

The experiments follow the ideation process of

agents solving the case study problem, with each

step’s outcome assessed against the original study.

The focus is not on comparing agent orchestration ar-

chitectures or prompt strategies.

Exploring LLM-based multi-agent systems may

signiﬁcantly enhance problem-solving tasks requir-

ing creative thinking, external resources, and broad

knowledge. Agent collaboration methods are likely

to become more investigated because of their poten-

tial to automate processes and optimize business op-

erations.

The paper is structured as follows. Section 2 de-

scribes the background and related work. Section 3

presents our TRIZ Agents framework. Section 4 de-

scribes our case study. Section 5 discusses the system

results and Section 6 summarizes our conclusions.

2 BACKGROUND

2.1 Related Work

Previous research has explored using LLMs to en-

hance the innovation process with TRIZ methodol-

ogy. However, the focus was more on creating work-

ﬂows for such models (Chen et al., 2024a), compar-

ing prompting strategies and LLM models (Jiang and

Luo, 2024). Previous works include building work-

ﬂows that strictly follow TRIZ steps and solve those

steps with reﬁned, prompt engineering strategies. Re-

sults showed a promising area of LLM usage - inno-

Figure 1: TRIZ Problem Solving Methodology.

vative problem-solving.

Studies have also suggested that agentic sys-

tems based on LLMs are signiﬁcantly effective (Wu

et al., 2023). With agentic GPT-3.5, it was possi-

ble to achieve results better than the foundation GPT-

4 model, a next-generation model (Hu et al., 2024).

This sparks a need to explore this approach.

Over the past few years, researchers and providers

have developed numerous models. OpenAI’s GPT

model is the most widely used, with many integra-

tions, which is why we decided to utilize it for our

experiments. Using other available LLMs, such as

Gemini, Llama, or Claude, would be equally justiﬁed.

This research does not focus on comparing LLMs.

Recent advancements in artiﬁcial intelligence, es-

pecially natural language processing, have signiﬁ-

cantly increased possible AI applications. The emer-

gence of Large Language Models and intensive re-

search in this area has shown that many new domains

can be supported with AI. Models, like OpenAI’s

GPT-3, have extensive knowledge about the world

(Atox and Clark, 2024). They have shown impressive

performance in passing tests concerning intelligence

and problem-solving abilities (Orr

u et al., 2023). Sur-

prisingly, LLMs have also demonstrated remarkable

capabilities in creative tasks, such as image gen-

eration (Oppenlaender, 2022), writing (Chakrabarty

et al., 2024), and even music generation (Yuan et al.,

2024). These results suggest using LLMs to solve

problems that require creative thinking, such as de-

signing innovations. Since current research suggests

that it is best to guide LLMs and carefully structure

their tasks (Minaee et al., 2024), the TRIZ methodol-

ogy seems perfect, with clear, progressive steps and

guidelines.

2.2 TRIZ Methodology

TRIZ is a well-known and broadly used method of

solving innovation problems. It was proposed by

Genrich Altshuller in the 1960s (Arciszewski, 2016).

It introduced a systematic approach to innovative

problem-solving. A critical insight from TRIZ cre-

ators is that problems and solutions reoccur across

domains of industries and sciences. The core idea

behind it is an assumption that there must be some-

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

197

thing in common, some pattern, in solving problems

described in thousands of patents. Four stages of

TRIZ are shown in Figure 1. This analysis allowed

researchers to deﬁne a step-by-step method for over-

coming innovation obstacles by abstracting them into

generalized patterns of invention. This allowed many

innovators to overcome seemingly impossible chal-

lenges without the need for unfavorable compromise.

Although TRIZ has primarily involved mechanical

problems, it was effectively transferred to other do-

mains, making it a universal method for innovative

problem-solving.

2.2.1 TRIZ Core Principles

The core principle of TRIZ is resolving contradictions

identiﬁed in the problem matter, where often, improv-

ing one feature implies worsening another. The TRIZ

methodology provides tools to support the resolution

of these contradictions. Two primary tools are the 40

Inventive Principles and the Contradiction Matrix.

The TRIZ process starts with a problem analy-

sis, resulting in the identiﬁcation of TRIZ parame-

ters. According to TRIZ methodology, 39 parame-

ters - such as Speed, Force, Temperature, etc. - rep-

resent system characteristics. By identifying the pa-

rameters relevant to the speciﬁc problem and their in-

terdependencies, the problem can be generalized and

structured within TRIZ.

The Contradiction Matrix is a tool for ﬁnding In-

novation Principles by providing two TRIZ parame-

ters: the improving feature and the worsening fea-

ture. The matrix contains information on which con-

tradictions can be solved by which principles. The

study of thousands of patterns enabled the structur-

ing of parameters and the creation of this tool. It is

worth noting that the contradiction matrix consists of

empty cells, meaning there are no inventive principles

for some contradictions.

The 40 Inventive Principles deﬁne ideas and ap-

proaches for resolving contradictions. They serve as

instructions or guidelines for overcoming given con-

ditions. They provide solutions to abstract problems,

which can be transformed into practical solutions in

the real world.

Figure 2: Agents collaboration methods (Wu et al., 2023).

2.2.2 TRIZ Complexity and Knowledge

Demands

TRIZ application follows clear guidelines; however,

to abstract problems effectively, identify contradic-

tions, and use TRIZ tools, it still requires deep in-

terdisciplinary knowledge of innovators (Moehrle,

2005). This makes it challenging to apply for non-

experts, and the ﬁnal result is highly dependent on

innovators’ expertise due to the high complexity of

the problems. In addition, following TRIZ steps like

analyzing system parameters, identifying contradic-

tions, or scrutiny of inventive principles is very labor-

intensive and time-consuming (Nassar and AbouR-

izk, 2016) (Cascini and Rissone, 2004). For this rea-

son, resulting in limited accessibility, TRIZ has been

adopted only in specialized innovation environments

like R&D centers or academic sites. Given its effec-

tiveness, it would be desirable to implement it in a

broader range of environments, especially those with-

out resources or access to experts.

2.3 Multi-Agent Systems and Language

Agents

Multi-Agent Systems (MAS), a subdomain of Dis-

tributed Artiﬁcial Intelligence (DAI), is a system that

consists of multiple autonomous entities known as

agents (Dorri et al., 2018). Each agent, put together

in a shared environment, can operate independently

but communicate and collaborate with others or in-

teract with the system to solve more complex tasks

(Van Harmelen et al., 2008). Agents alone can-

not perform complex tasks; however, after decom-

posing them into smaller subproblems, specialized

agents can solve them and, therefore, solve the pri-

mary task together. MAS are valued for ﬂexibility and

adaptability, especially in distributed intelligence and

decision-making (Serugendo et al., 2005). Figure 4

presents an exemplary architecture of an LLM-based

system, where tasks are distributed to multiple agents

equipped with context sources like tools or databases.

Those systems seem to have a signiﬁcant amount of

use cases because they, in a way, simulate the work of

human workers.

As opposed to traditional MAS that use for-

mal agent communication language protocols such as

FIPA ACL (Kone et al., 2000), our proposed approach

utilizes natural language prompts. Moreover, agents

communicate with one another by natural language.

While FIPA ACL provides a good, standardized, and

rigorous way for agent interactions, LLM-based com-

munication offers ﬂexibility and adaptability for com-

plex collaborative tasks.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

198

Figure 3: TRIZ Agents diagram, showing connections between key agents.

Multi-agent systems based on LLMs are currently

being researched extensively. Their core idea is to

handle complex problems through collaboration and

language abilities (Zhao et al., 2023). A notable ex-

ample of a multi-agent system is presented in the

study (Park et al., 2023), where agents demonstrate

human-like behavior within a shared interactive en-

vironment. This research underscores the poten-

tial of LLM-based multi-agent systems for simulat-

ing complex human communities. Studies have ex-

plored various implementations of teams, such as a

software engineering team (Chudziak and Cinkusz,

2024) (Cinkusz et al., 2025) and a multi-agent sys-

tem designed for big data analysis in ﬁnancial mar-

kets (Wawer et al., 2024). Other research focuses

on enhancing agents’ logical reasoning capabilities

and memory functions within LLM-based multi-agent

systems (Kostka and Chudziak, 2024).

2.3.1 Agent Orchestration

Along with research about the application of LLMs

in multi-agent systems, a new subdomain of those

studies has emerged called Agent Orchestration. It

explores how different agent architectures work and

produce results. In the past, there was no need for

that because most multi-agent systems were deter-

ministic microservices, and in general, the result of

given input could be known before running the sys-

tem. LLM agents introduced new areas of research

because of their stochasticity, ﬂexibility, and ability to

adapt or remember abstract information (Cheng et al.,

Figure 4: Architecture of LLM-based multi-agent system

enhanced with tools and data sources (H

andler, 2023).

2024). This promising ﬁeld of study involves design-

ing and evaluating agent orchestration architectures.

It includes investigating various ways agents inter-

act, such as supervised groups, equal collaboration,

and nested collaborating teams, as shown in Figure 2.

Agent orchestration techniques optimize the abilities

of each agent and, therefore, synergetically obtain the

best results.

Another related research ﬁeld is prompt engineer-

ing. This domain focuses on optimizing prompts to

achieve the best LLM outputs. Other studies demon-

strate that employing speciﬁc prompting strategies —

such as reasoning and acting (ReAct) (Yao et al.,

2023) or Chain-of-Thought Prompting (Wei et al.,

2023) — can signiﬁcantly enhance the quality of gen-

erated outcomes. The way in which roles and tasks

are described to an LLM agent is essential, and the

outcome strongly depends on it (Zamﬁrescu-Pereira

et al., 2023). It has to be clear what the structure of

the agent’s output should be. This study focuses on

basic prompt strategy, which consists of tasks, roles,

and contexts for the agent.

2.3.2 Agent Tools

What we know about agentic systems is that you can

never fully predict the outcome of an LLM’s actions.

While LLMs are pretty proﬁcient at planning and

decision-making, a default LLM model cannot per-

form actions or directly interact with the real world.

Tools enable these interactions by allowing LLMs to

generate function call outputs with speciﬁed argu-

ments. This mechanism can equip models with tools

such as web search, database queries, ﬁle reading and

writing, and more. Our goal is to provide agents with

tools to take action, apply grounding, and interact

with the system.

A known characteristic of LLMs is their tendency

to hallucinate, which comes from the fundamental

mathematical and logical structure of these models

(Banerjee et al., 2024). Although LLM knowledge

is broad and often accurate, it is not advisable to as-

sume that every generated text is truthful. This is es-

pecially important in our case, as we want agents to

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

199

reason based on fact-checked information. One way

to minimize the generation of incorrect knowledge is

by using a Retrieval-Augmented Generation (RAG)

mechanism (B

echard and Ayala, 2024). By incorpo-

rating RAG into a tool, agents can check necessary

knowledge by retrieving it from predeﬁned and reli-

able sources.

A simple web search is another helpful tool that

helps ground the models. It allows browsing the in-

ternet with queries about the topic. Eventually, the

agent gets valuable information, and based on it, it

generates a response.

2.4 Role of LLMs in Problem-Solving

and Innovation

Artiﬁcial intelligence has been applied to various and

numerous tasks, and TRIZ is one of them. Those im-

provements include estimation of ideality and level

of invention (Adams and Tate, 2009), or classiﬁca-

tion and analyzing engineering patents (Hall et al.,

2022). Recently, LLMs have opened a new set of

possible TRIZ enhancements. In recent works, re-

searchers have been exploring ways of using LLMs’

extensive broad knowledge and substantial reasoning

abilities to create systems supporting innovators in the

ideation process or entirely replace them (Chen et al.,

2024b)(Jiang and Luo, 2024). Those researchers have

uncovered the potential of LLMs in applications re-

quiring creative thinking. The limitations of these

methods become apparent with larger and more com-

plex tasks. Given that LLMs perform more effectively

on smaller subtasks (Chen et al., 2024a), exploring

Multi-Agent Systems has emerged as a natural so-

lution for handling such complexity. Such systems

could be ﬂexible and adaptable to the given problem,

enabling interdisciplinary problem-solving capabili-

ties.

3 TRIZ AGENTS

3.1 Workﬂow

As said before, the general idea simulates an innova-

tion team’s ideation process. The team is provided

with a problem description and, in return, must pro-

pose one or more solutions. The team works based

on the steps indicated in the case study paper. Each

step is done one by one, and a special agent docu-

ments each step. Documentation of the previous step

is provided to the team. One way to think about it

is that each step is a separate meeting where team

Figure 5: TRIZ Agents workﬂow steps and products.

members solve a problem. At the next meeting, team

members don’t remember what they said last time, but

they have documentation and the main ﬁndings of the

previous step. Based on that, they continue with the

following step. The process repeats until the last step

is achieved. After that, another documentation agent

compiles all documentation into one concise report

(Figure 7).

The ideal result would be that the team produces

the exact solutions with the same intermediate steps,

and the ﬁnal report’s content is the same as the case

study’s content. Each step will be compared and eval-

uated with the original case study. All steps are pre-

sented in Figure 5 for better workﬂow clarity. Docu-

menting the team’s progress at each step helps to track

the progress, but it also allows us to compare results

to the original case study’s corresponding parts.

3.1.1 Implementation

System implementation is based on LangGraph

(LangGraph, 2023), a framework designed to build

applications that integrate LLMs. They provide pre-

cise agent orchestration and control over the graph of

agents. Each agent was implemented as a node in

a graph, where edges are connections between them

- conditional or not. The agents’ tools are separate

nodes as well. This framework allows for more struc-

tured LLM text generation by using prompt templates

while calling LLM models. Messages of each step are

the primary context for LLM chat models to generate

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

200

Figure 6: Agent node with tool nodes.

answers. It is worth mentioning that in LangChain,

messages have classes, including Human or AI. In the

system, messages generated by models are converted

to human type so that the next model generates out-

puts as if it were conversing with a human.

For the LLM model, GPT-4o was selected due to

its high performance and relatively fast response time.

It supports function calls, which is necessary for tool

integration. Outputs of LLMs rely on the model tem-

perature parameter. In this study, this parameter was

arbitrarily set to 0.5 since it is a good compromise be-

tween reasonable and creative answers.

3.2 Agent Orchestration Architecture

For agent orchestration, a supervised team was cho-

sen. In the experiment, we call this agent a Project

Manager. This agent is responsible for following the

workﬂow, distributing tasks, and telling who should

act next. This agent, an orchestrator, manages the

team and makes sure it goes in the right direction. The

simple way to think about it is an actual group leader

in a real-life scenario, but its capabilities and respon-

sibilities are a bit narrow for the LLM efﬁcacy. Pro-

vided with workﬂow wrought on the case study pa-

per, Project Manager keeps track of the project and

steers the conversation so that the team solves tasks

one by one. Agents engage in the conversation once

they are asked to by Project Manager. They per-

form necessary actions, like tool usage, and provide

an appropriate answer. After that, attention is brought

back to Project Manager, who receives the message

and continues with management.

3.2.1 Team Composition

There are the following agents in the team:

Mechanical Engineer, Electrical Engineer,

Control Systems Engineer, Safety Engineer,

TRIZ Specialist, Operations Specialist,

Documentation Specialist and supervisor

Project Manager. Each agent’s role was deﬁned

based on case study analysis, which allowed to

specify what members the team would consist of.

3.2.2 Agent Proﬁling Prompts

Each agent has its own prompt specifying their per-

sona. It gives information about their ﬁeld of ex-

pertise, tools, and what their output should look like.

Each agent needs a proﬁling prompt, which will de-

ﬁne the persona behind it. Key elements that need

such prompts are:

• Name: Name of the agent. Necessary for ﬁnding

oneself in conversation.

• Role: Basic description of the agent’s place in

the team and its background for character devel-

opment.

• Tasks and Reponsibilities: Clearly states their

tasks and responsibilities so that LLM’s focus is

primarily on them.

• Context: Gives the agent an idea of what is the

current stage of the process. It can be messages of

the conversation or a summary of previous mes-

sages.

In addition, it is advisable to add some speciﬁc

instructions in the prompt that are critical for correct

system operation or just essential in the agent’s role.

The prompt template of Project Manager agent is

presented in Figure 8.

As it is shown, prompts are composed of system

messages and placeholders, such as messages in the

conversation or documentation on previous steps. It

is necessary to provide all inputs to the template in

order to run the model.

3.2.3 Tools of TRIZ Agents

In the real world, team members can use various tools

and knowledge bases, which is why we will provide

agents with the same. All agents except Project

Manager are equipped with tools. In the LangGraph

framework, agent nodes have conditional connections

with tool nodes since agents may or may not want to

use them. Contrarily, the edge from the tool node to

Figure 7: TRIZ Agents products.

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

201

Project Manager Prompt:

System Message

You are the Project Manager of a team

made up of different experts who are

working together to solve a complicated

engineering problem using the TRIZ

method.

...

Each step is solved and documented

separately. You are provided messages

of conversation of the current solving

step.

{messages}

System Message

You are also provided documentation of

previously performed steps, which tells

you about the progress:

{steps documentation}

System Message

Given the conversation above and the

documentation of previous steps, who

should act next? You have the following

decision options: {members names} or

FINISH.

...

Figure 8: Example of Project Manager prompt template.

Elements in {} brackets are inputs of prompt templates.

the agent node is non-conditional, so the right agent

gets the tool’s output. Figure 6 presents a worker

agent node with tool nodes.

The composition of the toolset for an agent de-

pends on their role in the team. All sets include a web

search tool, which is their elemental way of acquiring

knowledge. Since TRIZ Specialist is an important

team member, this agent is additionally equipped with

TRIZ-oriented tools such as:

• TRIZ Features Tool: Return list of 39 TRIZ pa-

rameters.

• Contradiction Matrix Tool: Returns worth apply-

ing inventive principles based on improving and

worsening features of the system.

• TRIZ Inventive Principles Tool: Tool returns

more details about given inventive principles.

• TRIZ RAG Tool: Returns answers to agent’s

queries based on TRIZ source materials, such as

books, websites, Wikipedia pages, etc.

4 CASE STUDY

To model team cooperation and evaluate the results

of the proposed agentic system, we conducted a case

study about the application of TRIZ for gantry crane

improvement based on (Luing et al., 2024). The au-

thors of this paper thoroughly describe each step of

their work. They strictly follow the TRIZ method-

ology. The ﬁnal outcome of the ideation process

consists of textual descriptions of the proposed ap-

proaches. During the problem analysis, authors iden-

tify system elements, perform CECA analysis, ﬁnd

physical and engineering contradictions, and apply

TRIZ principles using the TRIZ Matrix. In addition,

authors build graphs for analysis steps, which is again

feasible using LLMs. All those steps are a great usage

scenario for a team of agents. A team of predeﬁned

team members will cooperate and invent a solution for

the crane improvement problem.

It is worth mentioning the criteria for the chosen

case study. First of all, the problem description is suf-

ﬁciently characterized. All the intermediate steps and

solutions are clearly explained, which is vital during

the evaluation of results. Moreover, a case study was

published in the year 2024, which means that Chat

GPT-4o used in experiments was for sure not trained

on this text because the data cutoff for this model is

October 2023. This is extremely important since we

don’t want the model to know the solution - we want

the system to invent it.

5 RESULTS AND DISCUSSION

The proposed multi-agent system simulates a discus-

sion between team members as intended, resulting in

six documents about each workﬂow step and one ﬁ-

nal report, which compiles those documents. As an

input system message, the system receives a Human

type message with textual problems description of the

case study (Figure 9). It is copied from the case study

article not to omit information or add one.

It is essential to note the system’s outputs are not

deterministic, which comes from the nature of LLMs

(Ouyang et al., 2024). That is why the results de-

scribed here are generalized descriptions of how the

majority of runs went.

Another observation worth mentioning is the

number of iterations in the agent graph and the num-

ber of total tokens (prompt and output tokens). By

iterations in the graph, we mean the number of times

any node was called. This number varied from 60

to 80 calls. Total tokens of one runtime varied from

150,000 to 250,000.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

202

Input Human Message

Solve the following problem: Gantry cranes

find extensive application across various

industries, employed to move hefty loads

and dangerous substances within shipping

docks, building sites, steel plants,

storage facilities, and similar industrial

settings. The crane should move the load

fast without causing any unnecessary

excessive swing at the final position.

Moreover, gantry cranes which always lift

excessive load may result sudden stop of

the crane. The crane operators’ attempt

to lift heavier loads at a faster pace has

led to recurrent malfunctions, including

overheating, and the increased speed has

caused excessive swinging or swaying of the

lifted load, posing a safety hazard.

Figure 9: TRIZ Agents input message taken from case study

problem description.

5.1 Analysis of System’s Actions

Step 1: Deﬁning Engineering System. Project

Manager prompts Mechanical Engineer to identify

relevant elemnts of gantry crane engineering system.

This agent uses a web search tool to ﬁnd relevant in-

formation and, based on that, speciﬁes the engineer-

ing system. After that, Project Manager decides to

ﬁnish this step by documenting ﬁndings by Document

Specialist.

Step 2: Function Analysis. Project Manager

has at disposal the documentation from previous

step. Based on that, Project Manager this time

asks Mechanical Engineer to prepare the Func-

tion Analysis. The agent uses a web search tool

to generate an analysis. Next, a Control Systems

Engineer is asked to do the same from the control

systems engineering perspective. Like the previous

agent, it uses a web search tool to generate analysis.

Finally, Safety Engineer is asked to evaluate the

safety features and potential hazards of the system.

After that, Project Manager decides to move on to

the next step by documentation of Documentation

Specialist.

Step 3: Cause and Effect Chain Analysis (CECA).

Similar actions were taken by the agent while working

on Cause and Effect Chain Analysis. The Control

Systems Engineer was called to make the analysis

and used a web search tool. Then Safety Engineer

contributed to this step. The step ended with docu-

mentation on the Project Manager’s request. Part

of the conversation of step 3 was shown in Figure 10.

Step 4: Engineering Contradiction (EC) and Con-

tradiction Matrix. Project Manager asks only

TRIZSpecialist agent to do this task. The agent

ﬁrst checks the list of TRIZ features from which it

can choose. Based on that, it identiﬁes contradictions.

The agent uses another TRIZ tool, a contradiction ma-

trix, to ﬁnd inventive principles for those contradic-

tions. Next, it uses another tool that provides more

details about contradictions, which will help ﬁnd a so-

lution.

Step 5: Physical Contradiction. This step is also

solely performed by TRIZ Specialist by one an-

swer only without any tool use. With this information,

Project Manager ended this step with a documenta-

tion request.

Step 6: Solutions. Having all previous steps docu-

mented, Project Manager decided to begin the last

step, which was ﬁnding solutions based on identi-

ﬁed innovative principles. In this step, an agent who

contributed ﬁrst was Control Systems Engineer,

who used a web search tool to enhance its ﬁndings.

The following two agents who reviewed the proposed

solutions were Safety Engineer and Operations

Specialist. After solution propositions, Project

Manager directed the end of collaborative work. This

resulted in compiling all previously created docu-

ments into one ﬁnal report.

5.2 Assessment and Discussion

In this subsection, we will assess the actions per-

formed by the team. We will ﬁrst evaluate its perfor-

mance as a team and then compare it with solutions

proposed by the case study’s original research team.

Two separate runs of TRIZ Agents were chosen for

assessment, and their results were compiled.

5.2.1 TRIZ Agents Actions

Overall, all steps were executed, and in the end, the

team proposed solutions to the problem. There were

steps where the outcome was the result of interaction

between multiple agents (steps: 2, 3, 6), and there

were steps where only one agented contributed to the

result (steps: 1, 4, 5) - not counting the manager who

was only responsible for leading the discussion and

not ideation. How each step was solved and who con-

tributed to it was solely the decision of the Project

Manager. It is desirable that multiple agents cooper-

ate to solve each step because they may correct each

other’s outputs. This suggests adding more critique

instructions to Project Manager agent’s prompt for

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

203

ProjectManager

We are now moving to the Cause and

Effect Chain Analysis (CECA) step [...]

ControlSystemsEngineer, could you begin

by exploring the control-related issues

that might be contributing to these

problems? Please consider how the

control systems might be leading to

excessive speed or insufficient sway

control, and provide your insights.

ControlSystemEngineer

Calling web search tool: { "query":

"control systems for gantry cranes

excessive swinging overheating" }

Tool

[{ "url": "https:journals.sag ...,

"content": "Various attempts in

controlling gantry cranes ...

ControlSystemEngineer

Based on the information gathered [...]

we can identify several control-related

issues contributing to the problems of

excessive swinging and overheating in

gantry cranes. Here’s a detailed Cause

and Effect Chain Analysis (CECA) focusing

on these control aspects:

### Root Causes of Excessive Swinging and

Overheating

1. **Control System Sensitivity**: -

**Open-loop Control Systems**: These

systems are sensitive to ...

ProjectManager

SafetyEngineer, now that we have

identified the control-related root

causes of excessive swinging and

overheating, let’s explore the safety

implications of these issues. Please

provide your insights on ...

...

Figure 10: Example of messages in conversation between

agents during Step 3.

doubting the outputs of other agents. Another option

would be replacing each worker agent with a team of

agents performing one task. This would create a hi-

erarchical teams architecture, a promising but more

complex approach to agent orchestration.

Another observation is that the TRIZ Specialist

rarely used the RAG tool to request more information

about the TRIZ methodology. Although the agent’s

proﬁling prompt mentioned the RAG tool, it was not

explicitly stated that it should be utilized at every

generation, which may explain this behavior. TRIZ

Specialist agent was intentionally provided with

autonomy to decide whether to use external resources,

implicitly allowing it to rely on its internal knowl-

edge. Aim was to ensure that the agent follows the

TRIZ methodology closely, ideally using the context

of TRIZ literature. However, this autonomy caused

the agent to underuse the RAG tool, as the agent of-

ten generated responses without any tool usage or

with the usage of the web search tool. This outcome

demonstrates one of the challenges of prompt engi-

neering - balancing agent autonomy with consistent

obedience to certain assumptions about the agent. To

address this, future work will examine improving the

prompts to require or encourage using RAG tools.

5.2.2 Case Study Comparison

There are similarities between solutions, as shown by

comparing the result to the original case study’s solu-

tion. This section will focus on comparing those two

results.

Step 1: Deﬁning Engineering System. Most of

the Sub/System and System Components have been

found. In some, the name may be slightly different,

but the object is the same. Supersystems were not as

well identiﬁed. Results are compared in Table 1.

Table 1: Elements of the engineering system in the case

study: the upper row represents the system components,

while the lower row represents the supersystems. Elements

in bold were identiﬁed by TRIZ agents.

Gantry Legs, Wheel, Railway, Trolley, Trolley

Frame, Wire Rope, Motor, Hoist, Hook, Railway

Beam, Bridge Girder

Air particles/dust, Workers, Humidity, Thermal

Step 2: Function Analysis. In the case study, this

step was presented using a graph showing functions

as connections between elements. TRIZ Agents listed

the connections. Those connections labeled as useful

connections that overlap are: Trolley to Hoist (Motor),

Wire Rope to Hook. Connection Hoist to Wire Rope

was identiﬁed as useful, but case study researchers la-

bel it as insufﬁcient and harmful. On the other hand,

the connection Hoist to Load was identiﬁed as harm-

ful, which is comparable. The connection Environ-

mental Conditions to Load was successfully identi-

ﬁed, as the case study analysis highlights Dust, Hu-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

204

midity, and Thermal factors. However, six useful con-

nections present in the case study were not identiﬁed.

Additionally, the harmful connection between Worker

and System was missed by the TRIZ Agents.

Step 3: Cause and Effect Chain Analysis (CECA)

case study identiﬁed two main root causes, which are

the crane operator who does not follow standard op-

erational procedure by lifting the excessive weight of

load and fast speed of controlling the movement of the

cranes. Those causes were identiﬁed by TRIZ Agents

as Overloading and Rapid Movements. The system

identiﬁed ﬁve other root causes that are logical but do

not occur in the case study.

Step 4: Engineering Contradiction (EC) and Con-

tradiction Matrix. TRIZ Agents identiﬁed two en-

gineering contradicitons: Speed vs. Stability and

Load Capacity vs. Safety. This does not exactly

overlap with contradictions from a research paper.

However, TRIZ agents were not far from their solu-

tion because they identiﬁed Stability instead of Ob-

ject generated harmful factors and Load Capacity in-

stead of Weight of stationary object, which seem close

enough. Inventive principles were then found based

on those contradictions using the Contradiction Ma-

trix, which is a deterministic operation, which is why

there is no point in comparing those.

Step 5: Physical Contradiction. Physical con-

tradicitons were found exactly the same. Results are

presented in Table 2. In the next step, TRIZ Agents

derived only one, common with case study, inventive

principle. This may be due to the generally scarce

literature sources about physical contradictions iden-

tiﬁcation compared to engineering contradictions.

Table 2: Physical Contradictions identiﬁed by TRIZ

Agents, which overlap with case study.

Contradiction Contradictory Needs

Speed The crane must move quickly

to enhance productivity, but it

must also move slowly to pre-

vent load swinging.

Load

Capacity

The crane must lift heavy loads

to improve efﬁciency, but it

must lift lighter loads to ensure

safety and prevent overheating.

Step 6: Solutions Theoretical solutions proposed

were categorized based on the specializations of the

agents contributing to the ideation process. The ﬁrst

solution is the application of Sliding Mode Control

(SMC) with Antiswing Trajectory, which matches the

case study’s ﬁrst proposed solution. The solution

proposed by case study researchers intelligent circuit

breaker with essential sensors to monitor and regu-

late the current consumption was not found by TRIZ

Agents. During Step 6, the Electrical Engineering

agent was not involved in collaboration, which may

be the primary reason for this shortfall, as this agent

could have contributed the necessary idea in the elec-

trical domain. However, TRIZ Agents mention Ther-

mal Management Enhancements, which is in the area

of the last case study solution smart ventilation sys-

tem with features like self-cleaning ﬁlters and sealed

bearings.

6 CONCLUSIONS

6.1 Main Insights

The study showed that it is possible to model group

work on innovative problems using the TRIZ method-

ology with a multi-agent system based on LLMs. The

outputs of the TRIZ Agents’ steps were not identi-

cal to those in the case study in every detail, but each

step had commonalities. The solutions proposed by

TRIZ Agents, although different from the case study,

seemed logical. Of course, such multi-agent systems

are still far from achieving a complete and adequate

ideation process, but they can quickly produce so-

lutions or guide real-world researchers in a promis-

ing direction. Another insight is that the outcome

of the system strongly depends on agents’ proﬁling

prompts. Experimenting with many prompts eventu-

ally resulted in the team following the workﬂow.

6.2 Limitations

A substantial limitation of the proposed system is its

lack of a feedback loop in the teams’ actions. In the-

ory, it is possible, but the proﬁling prompt of Project

Manager does not mention doing it. The supervisor

agent never directed returning to any step to reﬁne re-

sults or try another approach. This is a severe draw-

back because every problem-solving process requires

iterating over ideas, revisiting steps, and rethinking

previous actions. The absence of such an essential

component is due to the high complexity of imple-

menting such a feature. For example, creating a loop

in agentic systems carries a signiﬁcant risk of causing

never-ending conversations between agents. This lim-

itation poses a promising research direction: making

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

205

multi-agent systems even more similar to real-world

teams.

There are also limitations connected with LLMs

properties like context window. In the discussed use

case, six workﬂow steps were outlined. In the case

of a workﬂow with more steps, the context window

could be too small to store all documentation from

previous steps. In fact, step documentation alone is a

way to handle this problem because instead of a long

list of messages, a summary is stored. This sparks a

need for long-time agent memory, where information

is available when necessary (Hatalis et al., 2023).

6.3 Future Work

Studies have shown a promising and worth-exploring

domain of multi-agent LLM systems. Examining how

other agent orchestration architectures might perform

would probably bring interesting insights. Addition-

ally, only the TRIZ methodology is considered in the

study. However, different methods, such as Lean In-

novation (Sehested and Sonnenberg, 2010) and De-

sign Thinking (Brown et al., 2008), are worth investi-

gating as they may work better with LLMs.

A promising direction for further research is to

compare how different prompt strategies function in a

multi-agent system scenario. Reﬁning prompts so that

agents iterate over ideas, repeat, and rethink previous

steps and actions could lead to substantial improve-

ments in the system. Although this research did not

focus on comparing prompts, it was surely a signiﬁ-

cant part of the implementation part because prompts

drastically inﬂuenced the system’s actions. Applying

an optimal prompt strategy might enhance the results.

Another interesting ﬁeld to investigate is cognitive

architectures in multi-agent systems. Cognitive archi-

tectures are augmented LLMs agents with the internal

ﬂow or external sources for tasks with grounding or

reasoning (Sumers et al., 2024). Building cognitive

agents could enable the system to learn problem de-

scriptions. There are already interesting studies about

using those concepts for enhancing decision-making

(Wu et al., 2024) and problem-solving (Sun, 2024).

REFERENCES

Adams, C. and Tate, D. (2009). Computer-aided triz ide-

ality and level of invention estimation using natu-

ral language processing and machine learning. In

Growth and Development of Computer-Aided Innova-

tion: Third IFIP WG 5.4 Working Conference, CAI

2009, Harbin, China, August 20-21, 2009. Proceed-

ings, pages 27–37. Springer.

Arciszewski, T. (2016). Inventive Engineering: Knowledge

and Skills for Creative Engineers. CRC Press, 1st edi-

tion.

Atox, N. and Clark, M. (2024). Evaluating large language

models through the lens of linguistic proﬁciency and

world knowledge: A comparative study. Authorea

Preprints.

Banerjee, S., Agarwal, A., and Singla, S. (2024). Llms will

always hallucinate, and we need to live with this.

Brown, T. et al. (2008). Design thinking. Harvard business

review, 86(6):84.

echard, P. and Ayala, O. M. (2024). Reducing hallucina-

tion in structured outputs via retrieval-augmented gen-

eration.

Cascini, G. and Rissone, P. (2004). Plastics design: inte-

grating triz creativity and semantic knowledge portals.

Journal of engineering design, 15(4):405–424.

Chakrabarty, T., Padmakumar, V., Brahman, F., and Mure-

san, S. (2024). Creativity support in the age of large

language models: An empirical study involving pro-

fessional writers. In Proceedings of the 16th Confer-

ence on Creativity & Cognition, pages 132–155.

Chen, D., Zhang, S., Zhuang, Y., Tang, S., Liu, Q., Wang,

H., and Xu, M. (2024a). Improving large models with

small models: Lower costs and better performance.

Chen, L., Song, Y., Ding, S., Sun, L., Childs, P., and Zuo,

H. (2024b). Triz-gpt: An llm-augmented method for

problem-solving.

Cheng, Y., Zhang, C., Zhang, Z., Meng, X., Hong, S., Li,

W., Wang, Z., Wang, Z., Yin, F., Zhao, J., and He, X.

(2024). Exploring large language model based intelli-

gent agents: Deﬁnitions, methods, and prospects.

Choinski, M. and Chudziak, J. A. (2009). Ontological

learning assistant for knowledge discovery and data

mining. In 2009 International Multiconference on

Computer Science and Information Technology, pages

147–155.

Chudziak, J. A. and Cinkusz, K. (2024). Towards llm-

augmented multiagent systems for agile software en-

gineering. In The 39th IEEE/ACM International Con-

ference on Automated Software Engineering (ASE

2024), Sacramento, CA, USA.

Cinkusz, K., Chudziak, J. A., and Niewiadomska-

Szynkiewicz, E. (2025). Cognitive agents powered by

large language models for agile software project man-

agement. Electronics, 14(1).

Czinki, A. and Hentschel, C. (2016). Solving complex prob-

lems and triz. Procedia Cirp, 39:27–32.

Dorri, A., Kanhere, S. S., and Jurdak, R. (2018). Multi-

agent systems: A survey. Ieee Access, 6:28573–

28593.

Hall, S., Mollan, C., Pandey, V., and Mourelatos, Z. (2022).

Triz mapping and novelty detection of engineering de-

sign patents using machine learning. In International

Design Engineering Technical Conferences and Com-

puters and Information in Engineering Conference,

volume 86267, page V006T06A044. American Soci-

ety of Mechanical Engineers.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

206

andler, T. (2023). A taxonomy for autonomous llm-

powered multi-agent architectures. In KMIS, pages

85–98.

Hatalis, K., Christou, D., Myers, J., Jones, S., Lambert, K.,

Amos-Binks, A., Dannenhauer, Z., and Dannenhauer,

D. (2023). Memory matters: The need to improve

long-term memory in llm-agents. In Proceedings of

the AAAI Symposium Series, volume 2, pages 277–

280.

Hu, S., Lu, C., and Clune, J. (2024). Automated design of

agentic systems.

Ilevbare, I. M., Probert, D., and Phaal, R. (2013). A re-

view of triz, and its beneﬁts and challenges in practice.

Technovation, 33(2-3):30–37.

Jiang, S. and Luo, J. (2024). Autotriz: Artiﬁcial ideation

with triz and large language models.

Kone, M. T., Shimazu, A., and Nakajima, T. (2000). The

state of the art in agent communication languages.

Knowledge and Information Systems, 2:259–284.

Kostka, A. and Chudziak, J. A. (2024). Synergizing logical

reasoning, long-term memory, and collaborative intel-

ligence in multi-agent llm systems. In Paciﬁc Asia

Conference on Language, Information and Computa-

tion (PACLIC 38), Tokyo, Japan. In press.

LangGraph (2023). LangGraph. https://langchain-ai.

github.io/langgraph/. Accessed: 10-01-2025.

Loh, H. T., He, C., and Shen, L. (2006). Automatic classiﬁ-

cation of patent documents for triz users. World Patent

Information, 28(1):6–13.

Luing, N. S. S., Toh, G. G., and Chau, G. H. (2024). Appli-

cation of triz for gantry crane improvement. In Jour-

nal of Physics: Conference Series, volume 2772, page

012004. IOP Publishing.

Mahto, D. (2013). Concepts, tools and techniques of prob-

lem solving through triz: A review. International

Journal of Innovative Research in Science, Engineer-

ing and Technology, 2(7).

Minaee, S., Mikolov, T., Nikzad, N., Chenaghlu, M.,

Socher, R., Amatriain, X., and Gao, J. (2024). Large

language models: A survey.

Moehrle, M. G. (2005). What is triz? from conceptual ba-

sics to a framework for research. Creativity and inno-

vation management, 14(1):3–13.

Nassar, N. and AbouRizk, S. (2016). Introduction to tech-

niques for resolving project performance contradic-

tions. Journal of Construction Engineering and Man-

agement, 142(8):04016027.

Oppenlaender, J. (2022). The creativity of text-to-image

generation. In Proceedings of the 25th international

academic mindtrek conference, pages 192–202.

Orloff, M. A. (2006). TRIZ. Springer.

Orr

u, G., Piarulli, A., Conversano, C., and Gemignani, A.

(2023). Human-like problem-solving abilities in large

language models using chatgpt. Frontiers in artiﬁcial

intelligence, 6:1199350.

Ouyang, S., Zhang, J. M., Harman, M., and Wang, M.

(2024). An empirical study of the non-determinism

of chatgpt in code generation. ACM Transactions on

Software Engineering and Methodology.

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

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

Interactive simulacra of human behavior.

Sehested, C. and Sonnenberg, H. (2010). Lean innovation:

a fast path from knowledge to value. Springer Science

& Business Media.

Serugendo, G. D. M., Gleizes, M.-P., and Karageorgos, A.

(2005). Self-organization in multi-agent systems. The

Knowledge engineering review, 20(2):165–189.

Sumers, T. R., Yao, S., Narasimhan, K., and Grifﬁths, T. L.

(2024). Cognitive architectures for language agents.

Sun, R. (2024). Can a cognitive architecture fundamen-

tally enhance llms? or vice versa? arXiv preprint

arXiv:2401.10444.

Van Harmelen, F., Lifschitz, V., and Porter, B. (2008).

Handbook of knowledge representation. Elsevier.

Wawer, M., Chudziak, J. A., and Niewiadomska-

Szynkiewicz, E. (2024). Large language models and

the elliott wave principle: A multi-agent deep learn-

ing approach to big data analysis in ﬁnancial markets.

Applied Sciences, 14(24).

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B.,

Xia, F., Chi, E., Le, Q., and Zhou, D. (2023). Chain-

of-thought prompting elicits reasoning in large lan-

guage models.

Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang,

L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H.,

White, R. W., Burger, D., and Wang, C. (2023). Au-

togen: Enabling next-gen llm applications via multi-

agent conversation.

Wu, S., Oltramari, A., Francis, J., Giles, C. L., and Rit-

ter, F. E. (2024). Cognitive llms: Toward human-

like artiﬁcial intelligence by integrating cognitive ar-

chitectures and large language models for manufactur-

ing decision-making. Neurosymbolic Artiﬁcial Intelli-

gence.

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan,

K., and Cao, Y. (2023). React: Synergizing reasoning

and acting in language models.

Yuan, R., Lin, H., Wang, Y., Tian, Z., Wu, S., Shen, T.,

Zhang, G., Wu, Y., Liu, C., Zhou, Z., Ma, Z., Xue, L.,

Wang, Z., Liu, Q., Zheng, T., Li, Y., Ma, Y., Liang, Y.,

Chi, X., Liu, R., Wang, Z., Li, P., Wu, J., Lin, C., Liu,

Q., Jiang, T., Huang, W., Chen, W., Benetos, E., Fu,

J., Xia, G., Dannenberg, R., Xue, W., Kang, S., and

Guo, Y. (2024). Chatmusician: Understanding and

generating music intrinsically with llm.

Zamﬁrescu-Pereira, J., Wong, R. Y., Hartmann, B., and

Yang, Q. (2023). Why johnny can’t prompt: how

non-ai experts try (and fail) to design llm prompts. In

Proceedings of the 2023 CHI Conference on Human

Factors in Computing Systems, pages 1–21.

Zhao, Z., Lee, W. S., and Hsu, D. (2023). Large language

models as commonsense knowledge for large-scale

task planning. In Oh, A., Naumann, T., Globerson,

A., Saenko, K., Hardt, M., and Levine, S., editors,

Advances in Neural Information Processing Systems,

volume 36, pages 31967–31987. Curran Associates,

Inc.

TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

207