AgentFlow: A Context Aware Multi-Agent Framework for

Dynamic Agent Collaboration

Gayathri Nettem

†

, Disha M.

†

, Aavish Gilbert J.

†

, Skanda Shreesha Prasad

†

and S. Natarajan

PES University, Department of Computer Science and Engineering, India

Keywords: Multi-Agent Systems, Dynamic Agent Selection, Agent Collaboration, Context Awareness,

Conversational Agents, Tools.

Abstract: Multi-agent systems have long been recognized for their potential in solving complex problems. This paper

presents a new framework that focuses on context-awareness and adaptability to tasks dynamically. Unlike

traditional agentic approaches, our method involves multiple specialized agents working together, each guided

by a set of strategies. These agents dynamically switch roles, utilize workflows to organize task progression,

and leverage the perception loop to ensure context-informed decisions and seamless collaboration. The result

is a system that consistently produces more accurate, coherent, and creative responses across a variety of

domains. Empirical evaluations using benchmarks like HumanEval and MMLU show substantial

improvements over single-agent and multi-agent systems.

1 INTRODUCTION

The field of artificial intelligence has witnessed a

paradigm shift with the advent of Large Language

Models (LLMs), which have demonstrated

remarkable capabilities across a wide range of tasks.

Concurrently, multi-agent systems (MAS) have long

been recognized for their potential in solving

complex, distributed problems. This paper introduces

a novel framework that synergizes these two powerful

concepts, presenting a multi-agent system composed

of LLM-based agents.

Traditional MAS frameworks, while effective in

many scenarios, often lack the flexibility and

contextual understanding that LLMs can provide.

Moreover, existing LLM applications typically

operate as single-agent systems, limiting their ability

to tackle complex, multi-faceted problems that

require coordinated efforts. Our proposed framework

addresses these limitations by creating a dynamic,

adaptive system of LLM-based agents.

The core innovation of our framework lies in its

ability to leverage the context, be it through the

https://orcid.org/0009-0008-1458-8702

https://orcid.org/0009-0000-5794-385X

https://orcid.org/0009-0003-1878-902X

conversation history or any specialized information

that is not one of the parameters of the LLMs.

Conversation history in our framework is stored in a

specialized database from which agents can fetch

data. It is important to note that it would serve as the

context for the LLMs. This would help in ensuring all

LLMs are aware of the conversation, and do not

hallucinate. Context-Awareness can also be

introduced using tools. The tools available can be

used to fetch data from a database, or even the internet

enabling agents to work with any proprietary data that

is not a part of an LLMs parameters.

Another core feature of our framework is its

ability to choose the next agent dynamically. Taking

into account the dialogue archive, the stage of the task

completed and the abilities of each agent, the

framework intelligently determines which agent gets

the chance next. This process enables the framework

to introduce collaboration, where all agents work

together for a common goal.

Through this work, we aim to contribute to a

significant advancement in the areas of natural

language processing and artificial intelligence,

https://orcid.org/0009-0002-9448-2292

https://orcid.org/0000-0002-8689-5137

†Equal Contribution and Primary Authors

Nettem, G., Disha, M., J., A. G., Prasad, S. S. and Natarajan, S.

AgentFlow: A Context Aware Multi-Agent Framework for Dynamic Agent Collaboration.

DOI: 10.5220/0013375700003890

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 687-693

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

687

offering a scalable, secure, and context-aware

foundation for developing sophisticated multi-agent

LLM systems.

2 RELATED WORK AND

BACKGROUND

The intersection of Large Language Models (LLMs)

and multi-agent systems has become an interesting

area of research, driven by the potential to create

complex, intelligent, and adaptable systems. Existing

work in this field has laid a strong foundation,

exploring various aspects of multi-agent LLM

systems.

A substantial amount of work has been done in

extending the functionality of MAS by incorporating

LLMs as part of the system. For example,

Chan, et al.

(2023)

developed a framework by which agents use

LLMs to formulate plans in natural language and log

persuasive messages.

Zhang, Y., et al. (2023) also used

LLMs which help the agents learn from others and

make improvements in case of evolving

surroundings. They have shown that LLMs can

contribute to enhancing agreements, cooperation and

conflict resolution among the agents.

Another important feature of many multi-agent

systems is dynamicity.

Lin, T., et al. (2023) described a

system by which agents can be deployed

differentiated by a task or a role.

Chen, Y., et al. (2021)

examined the issues and advantages of dynamic

collaboration of agents using flexibility and

adaptability factors.

Another interesting issue in LLM-based multi-

agent systems is security. A multi-agent system is

prone to jailbreaking. In recent works, studies have

focused on finding ways to counteract jailbreaking

and other malicious use of the product.

Su, X., et al.

(2022)

proposed a security model that relies on LLMs

to survive adversarial attacks and identify its

presence.

Wang, Y., et al. (2024) explored the

successful training of agents to reduce their level of

compliance using reinforcement learning. These

efforts emphasise that strong precautions are essential

to guarantee the efficient functioning of LLM-based

multi-agent systems and the credibility of the agents

involved in the systems. While we highlight the

importance of security, it is currently not a scope of

our project, but it can be integrated with our

framework.

One area that is crucial to intelligent agents is

context, and its understanding is vital for agents’

performance in complex conditions.

Liu, X., et al.

(2020) presented a framework in which agents apply

LLMs for comprehending useful information from an

environment and making decisions. Another study

further emphasized the importance of context for

agents which was done by Nguyen, H., et al. (2022).

These works raise the question of whether the context

of agents should be sufficiently flexible to allow the

desired goals to be attained.

2.1 Our Contribution

Building upon the work done in other frameworks,

our research aims to contribute to the field of LLM-

based multi-agent systems by developing a

framework that addresses the following key

challenges:

1. Context Awareness: To address this, we look

at how contextual awareness of the

environment can be achieved using LLMs thus

allowing agents to make informed decisions in

its interaction with the environment.

2. Dynamic Agent Management: The proposed

framework facilitates flexible and cooperative

agent interactions.

2.1.1 Context Awareness

This is accomplished by creating agents that have

access to every agent's reaction and thought process

in the workflow, compiled in an agent-understandable

format in terms of intra- and inter-agent memory.

This enables the agent to make well-informed

choices.

Most agentic systems provide memory to agents

by passing in a summary of the previous responses by

the other agents, but this method is inherently flawed

as it fails to describe the most crucial aspect of

communication. That is the order chronology of

events by the other agents during the conversation.

Context Awareness is also achieved through the

usage of tools, through which the agents will have

access to real time data instead of relying on the

training data from LLMs, This inturn reduces

hallucinations as references are made to an external

reliable data source.

2.1.2 Dynamic Agent Management

We introduce the concept of workflows and

perception loop into multi agent systems. The task at

hand is divided into smaller workflows, which are

then included into the perception loop to enable

dynamic switching between them.

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

688

Workflows are made to iterate repeatedly until the

task is completed satisfactorily, at which point it

decides which process is the best to proceed to next.

This gives the agents the freedom to work together

and solve problems one at a time rather than getting

disoriented.

3 PROPOSED ARCHITECTURE

In our solution, we have selected an approach where

separate tasks will have distinct workflows, all of

which will be managed utilizing a layered

architecture approach.

3.1 Layered Architecture

The proposed architecture for our system is designed

to address the limitations of existing methods The

modules are structured in the following manner:

1. Agent: It acts as a building block of the

framework. Each specialized agent is

extended from the Base Agent class.

2. Memory: Maintains all the agent and tool

interactions within an agent and with

external agents.

3. Perception Loop: Layer responsible to

ensure the dynamic communication is intact

and consensus is achieved.

4. Tools: Access to external APIs, SDKs, etc

to enhance agent functionality.

3.2 High-Level Workflow

The high-level workflow of the system can be

summarized in the following steps:

1. The agents in the framework get initialized

and invoked with suitable arguments like

agent prompt template, agent blueprint, json

template, task, tools, etc.

2. These agents are placed in the perception

loop, which combines workflows that are

independent to the agent. It essentially steers

the agent in a specific route while taking into

account a number of factors, including the

agent's present goal in a multi-agent

conversation, the type of decisions that need

to be made qualitative or quantitative,

relevant tool use, etc.

3. Throughout the perception there could be

multiple workflows that agents will use as a

guidance mechanism.

4. The agents generate responses and update

the memory of the system.

5. Both information regarding intra-agent and

inter-agent conversations along with tools

usage information is logged in so the next

agent that comes up can make a more

contextually aware decision.

6. The workflows themselves can loop in

multiple iterations, refer to other workflows

for guidance, and finally arrive at a common

consensus.

3.3 Sample Workflows

3.3.1 Workflow for Coding Questions

1. The Strategist generates a strategy based on

the user query. It helps in dividing a complex

problem into multiple simple sub-problems.

Figure 1: Coding Agents.

2. The coder agent codes according to the

strategy.

3. The code is then reviewed by the reviewer

agent. The reviewer agent checks the

correctness of code against edge cases.

4. The router agent routes the flow to the final

reporter agent only if the code passes the

review. If the code is incorrect, then the flow

is routed to the coder agent.

3.3.2 Workflow for Research Questions

1. The Strategist generates a strategy and a

search term based on the user query.

2. The web search tool returns a list of useful

urls, relevant to the user query.

3. The selector agent chooses one or more urls

to scrape information.

4. The reporter agent summarizes the

information fetched by the web scraper tool.

5. Reporter's response is reviewed by the

reviewer agent. If the response is both

accurate and comprehensive, the flow

terminates.

AgentFlow: A Context Aware Multi-Agent Framework for Dynamic Agent Collaboration

689

6. If the response does not pass the review, then

the router agent routes the flow to the most

suitable agent.

Figure 2: Research Agents.

3.3.3 Workflow for Specialized Agents

This proposed framework allows for the selection of

any Large Language Model (LLM) of the user’s

choice. This makes the model independent of the

computational needs and preferences and thus

versatile for use.

Figure 3: Specialized Agents.

The agent selection component is a key part of the

framework and is specifically meant for targeting

particular niches of application with suitable classes

of agents. A number of categories are available –

research, coding and specialized agents that are more

specific. It also presents the flexibility of a modular

design in any computer system by assigning the most

appropriate and efficient agent for each

computational task. Most notable is the increased

competence arising from the application of domain –

specialized knowledge and skills as offered by the

specialized agents.

The idea of context awareness is supported with

the help of the web search tools, which is considered

as one of the primary helpers in the process of

contextual development of the whole system.

Through executing organic searches and selecting

the most relevant links only, the tools offer the

dynamic mechanism for information retrieval on the

basis of the context awareness mechanism. The

context awareness mechanism transmutes the system

from a simple information processor into a highly

effective intelligent platform, which is able to offer

the detailed contextualized information concerning

the selected queries.

4 EXPERIMENTAL SETUP AND

EVALUATION METRICS

4.1 Dataset and Evaluation Metrics

To comprehensively evaluate the performance of our

proposed framework, we employed a combination of

existing datasets and metrics.

1. HumanEval(Chen, M., et al., 2021): A

benchmark for evaluating code generation

capabilities, focusing on tasks that require

understanding natural language instructions

and generating correct code.

2. MMLU(Hendrycks, D., et al., 2021): A

multi-task language understanding

benchmark designed to assess an LLM's

ability to perform well on a variety of tasks,

including reading comprehension, question

answering, and summarization.

4.2 Performance Metrics

1. Accuracy: The number of correct answers

given by the framework relative to the total

number of questions asked.

2. Number of API Calls: This metric tracks

down the number of API calls made by the

framework to retrieve the final response.

3. Response Time: The time taken from

initiating an API call to fetching the final

response. Lower response time is critical for

real-time applications like chatbots.

However since other frameworks do not use

response time as a benchmark for

evaluation, we have decided to not include

this metric in our analysis.

5 RESULTS

We have evaluated our framework using two

universally accepted evaluation metrics, namely,

MMLU and HumanEval. We have obtained

promising results proving that our framework

performs better than almost all of the Single Agent

LLMs available.

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

690

Figure 4: An illustration depicting an example of how coding agents communicate.

Figure 5: An illustration demonstrating the workflows in a perception loop.

AgentFlow: A Context Aware Multi-Agent Framework for Dynamic Agent Collaboration

691

Table 1:

HumanEval

zero-shot performance.

Model HumanEval Score

CodeGeeX(Zheng, Q., et

al.,2023)

18.9

Claude-instant-1 31.1

CodeGen-Mono 32.9

SteloCoder

(

Pan, J., et al., 2023

)

34.1

PaLM Coder 43.9

CodeX

(

Chen, M., et al., 2021

)

47.0

GPT-3.5-turbo 57.3

GPT-4-turbo 57.9

CodeX + CodeT 65.8

Claude 3 Sonnet 73.0

Gemini Ultra 74.4

Claude 3 Haiku 75.9

AgentFlow 84.14

When tested with Zero-Shot HumanEval

Benchmark, our agent system achieved a remarkable

score of 84.14. Our agent system not only exceeds the

capabilities of the State of the Art Single LLM

models, but does so by a significant margin. The State

of the Art Single LLMs are often constrained by static

reasoning pathways, while a Multi-Agent System like

ours provides versatility in the control flow.

Figure 6: Accuracy (%) on the MMLU dataset.

Our specialized agents were tested across

different categories of MMLU dataset( fig:Accuracy

on MMLU dataset). The framework's performance

varies by subject, with biology achieving the best

accuracy (84.02%), closely followed by journalism

(78.75%) and financial analysis (81.09 %). It

performs slightly lower in Cyber Security (67%) and

Statistics (68.52%), but it retains high accuracy in

STEM subjects like Astronomy (73.02%) and

Physics (76.59%). These findings suggest that while

our methodology is especially useful in domains that

demand a high level of analytical expertise, including

biology and financial analysis, it might be

strengthened in domains like cybersecurity.

5.1 Comparison of AgentFlow and

Other Frameworks

In comparing AGENTFLOW with other frameworks

- LLM Blender (

Jiang, D., Ren, X., & Lin, B. Y. (2023)

LLM Debate, DYLAN- AGENTFLOW consistently

achieves superior performance, particularly in

Humanities (67.08%), STEM (70.21%), and Other

categories (75.75%). In STEM, AGENTFLOW

outperforms all frameworks, showcasing a well-

rounded capability for scientific and technical tasks.

This advantage suggests AGENTFLOW's

exceptional adaptability and contextual

understanding in multi-disciplinary problem-solving,

aligning with its designed architecture for dynamic

collaboration and reasoning. Additionally,

AGENTFLOW's balanced performance across Social

Sciences (74.07%) suggests it holds robust

interpretative abilities across both qualitative and

quantitative domains, highlighting its utility in

socially focused analyses.

6 FINDINGS AND FUTURE

DIRECTION

The development of the Large Language Model

(LLM) based on multi-agent systems is a relevant

step forward in computational intelligence and

competence, differentiating it from conventional

single-agent systems. During the preliminary studies,

the system has shown an ability to produce responses

of an impressively high complexity level relevant to

the context in which it is used. Agents use context

awareness skills to federate context specific methods,

and as a result agents are capable of generating

accurate outputs that meet the complexity of the

computational problems posed to the agents.

Table 2: Comparison of scores across different frameworks.

Category Single

Execution

LLM Blender LLM Debate DYLAN AGENTFLOW

Humanities 59.8 60.4 59.8 62.1 67.08

STEM 62.9 66.3 69 69.7 70.21

Social Science 74 75.2 77.4 79.1 74.07

Other 71.8 70.7 75.5 75.5 75.75

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

692

Future research will include enhancing evaluation

metrics. We have noticed that the evaluation metrics

available truly do not quantify the abilities of Multi-

Agent systems, and thus, a novel, industry standard

evaluation metric and dataset is needed to fulfil this

gap. Further enhancements to the framework could

include dynamic agent creation, wherein, the

framework’s built-in agents of the system create

specialized agents based on the need.

7 CONCLUSION

This paper has introduced a novel framework for

building robust and adaptive LLM-based multi-agent

systems. Our strategy embeds dynamic agent

selection, collaboration of agents, and an entire suite

of tools to enhance agent capabilities and contextual

awareness. Results show this framework to be

successful in producing more nuanced, coherent, and

creative answers than in traditional LLMs. This paves

the way for a significant advancement in multi-agent

systems, particularly in domains requiring nuanced

understanding, adaptability, and security.

REFERENCES

Chan, C.-M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S.,

Fu, J., & Liu, Z. (2023). ChatEval: Towards better

LLM-based evaluators through multi-agent debate.

arXiv e-prints, arXiv:2308.07201. https://doi.org/10.

48550/arXiv.2308.07201

Zhang, Y., et al. (2023). Language model-based multi-agent

systems: A survey. arXiv preprint, arXiv:2303.16136.

https://doi.org/10.48550/arXiv. 2303.16136

Su, X., et al. (2022). Learning to cooperate in multi-agent

systems with language models. arXiv preprint,

arXiv:2205.14051. https://doi.org/10.48550/arXiv.

2205.14051

Wang, Y., et al. (2024). Securing LLM-based multi-agent

systems: A survey. arXiv preprint, arXiv:2402.01234.

https://doi.org/10.48550/arXiv. 2402.01234

Lin, T., et al. (2023). Reinforcement learning for ethical and

secure LLM-based multi-agent systems. arXiv preprint,

arXiv:2309.08765. https://doi.org/10.48 550/arXiv.

2309.08765

Chen, Y., et al. (2021). Dynamic agent creation and

management in LLM-based multi-agent systems. arXiv

preprint, arXiv:2112.04567. https://doi.org/10.48550/

arXiv.2112.04567

Liu, X., et al. (2020). Challenges and opportunities in

dynamic agent collaboration. arXiv preprint,

arXiv:2009.02345. https://doi.org/10.48550/arXiv.

2009.02345

Nguyen, H., et al. (2022). Context-aware LLM-based

agents: A survey. arXiv preprint, arXiv:2207.07890.

https://doi.org/10.48550/arXiv.2207.07890, https://doi.

org/10.48550/arXiv.2204.05999

Zheng, Q., et al. (2023). Codegeex: A pre-trained model for

code generation with multilingual benchmarking on

humaneval-x. In Proceedings of the 29th ACM

SIGKDD Conference on Knowledge Discovery and

Data Mining (pp. 2369–2378).

Pan, J., et al. (2023). SteloCoder: A decoder-only LLM for

multi-language to Python code translation. arXiv

preprint, arXiv:2310.01234. https://doi.org/10.

48550/arXiv.2310.01234

Nijkamp, E., et al. (2023). CodeGen: An open large

language model for code with multi-turn program

synthesis. In Proceedings of the Eleventh International

Conf. on Learning Representations.

Jiang, D., Ren, X., & Lin, B. Y. (2023). LLM-Blender:

Ensembling large language models with pairwise

ranking and generative fusion. In Proceedings of the

61st Annual Meeting of the Association for

Computational Linguistics.

Chen, M., et al. (2021). Evaluating large language models

trained on code. arXiv preprint, arXiv:2107.

https://doi.org/10.48550/arXiv.2107

Hendrycks, D., et al. (2021). Measuring massive multitask

language understanding. In Proceedings of the

International Conference on Learning Representations.

AgentFlow: A Context Aware Multi-Agent Framework for Dynamic Agent Collaboration

693