Agent studio: A multi-agent system for systems engineering

The following blog explores an emerging area of AI – the multi-agent system.  It is important to point out that is a proof-of-concept implementation designed to showcase the future of engineering.

Achieving a state of flow, often referred to as “being in the zone,” occurs when an individual is fully immersed in an activity and experiences effortless focus. When this state is experienced by all members of a team simultaneously, it is known as team flow. In this scenario, team members work together seamlessly, exhibiting synchronization, organization and collaborative progress on a project.

However, reaching team flow is not a matter of chance; it requires strategic planning. This includes prioritizing shared goals and minimizing distractions. Managers and their teams often face unpredictable and unexpected challenges that can disrupt their effectiveness and threaten project deadlines. Even with careful precautions and planning, it is unrealistic to expect that team flow can be maintained consistently over time.

Scenario: Urgent request for Systems Engineering

Sarah is a system engineer who manages a team that includes the following members:

• Jim, Project Manager
• Laura, System Architect
• Omar, Requirements Engineer
• Dorlis, Engineering Assistant

An urgent request comes in: Sarah and her team must investigate hybrid vehicle (HV) designs that could achieve a range of 560km and a speed of 110km/h within 10 seconds. This task is impossible to fit into their already busy day without breaking the team’s flow and rearranging their priorities and commitments.

Achieving and maintaining that team flow is critical for Sarah’s team. This state of focused productivity allows them to efficiently meet their objectives without distractions. They need to stay fully focused on their goals. When the team is in flow, they are unstoppable, continuously working towards their objectives without interruption.

Using standard AI tools

When a team encounters an urgent task in addition to their existing workload, Generative AI (GenAI) and large language models (LLMs) can assist by automating or accelerating various workflows. When leveraged properly, AI is a powerful tool to enhance productivity and help maintain team flow.

When integrated with Simcenter engineering tools, simulation users gain a novel way to interact with their modeling and simulation environment through a natural language interface. Tedious tasks that previously required numerous clicks and specialized expertise can now be accomplished simply by asking a question, further supporting the team’s ability to stay in flow.

Sarah and her team have recognized the benefits of this approach. However, like many users, they quickly realized that workflow improvements are necessary when tasks become more complex, such as investigating HV designs to achieve specific performance targets. Even with effective prompt engineering techniques, prompts can become excessively long, making it difficult for LLM to understand and follow all the instructions. Furthermore, real-world tasks like systems engineering often require the chatbot to utilize tools, such as internal design documents and knowledge base through Retrieval Augmented Generation (RAG), simulation tools like Simcenter Amesim and architecture exploration tools like Simcenter Studio Efficiently integrating these tools with an LLM presents a significant challenge but is essential for maintaining the team’s flow.

A new era of multi-agent systems

This is where agentic workflows and multi-agent systems come into play.  A multi-agent system consists of multiple interacting intelligent agents, which can be software, robots, or autonomous entities acting independently within an environment. Key characteristics include:

• Autonomy
• Social ability for agent interaction
• Reactivity to environmental changes
• Pro-activeness in fulfilling objectives

Multi-agent systems offer scalability benefits by adding agents, robustness where one agent’s failure doesn’t impact the system, and flexibility for dynamic tasks and environmental adaptation, making them an appealing paradigm for systems requiring collaborative autonomy to achieve collective goals. They are incredibly effective in tackling complex tasks while still providing a straightforward, intuitive natural language interface.

Break down the problem

Imagine a multi-agent system designed for systems engineering tasks, where each agent specializes in a specific role, such as system architect, requirements engineer, and so on. This setup mirrors the dynamics and expertise that enhance Sarah and her team’s effectiveness. Each agent is equipped with the necessary capabilities, data, and understanding of the objectives. Rather than operating as standalone tools, AI agents can experience “team flow” and integrate seamlessly to boost the team’s productivity and workflow.

For instance, in response to Sarah’s HV design request, an eight-agent architecture Agent Studio has been proven to significantly increase automation levels compared to a traditional chatbot workflow.

The engineering assistant agent has access to the Simcenter knowledge graph, preloaded with all engineering best practices, historical models, and libraries. Based on Sarah’s request, this agent identifies a list of modeling artifacts (e.g., vehicle, battery, axle, gearbox components) and passes them to the architect agent.

The architect agent, trained to write in ACEL (Architecture-centric Exploration Language), a domain-specific language (DSL) of Simcenter Studio, receives this list of artifacts and creates several feasible EV architectures. These iterations are validated through Simcenter Amesim simulations, resulting in a few architectures and their corresponding simulation results. These results are then forwarded to the requirements agent.

The requirements agent, trained to write in SysMLv2 (a standard modeling language for validating designs with formally defined requirements), formulates Sarah’s performance requirements for the vehicle in SysMLv2. The architectures are then validated against this requirement in Simcenter Studio.

The most suitable architecture is then identified and sent to the project manager agent, who reports the results to Sarah.

When critical decisions need to be made by Sarah, the multi-agent system will wait for Sarah’s feedback before proceeding. Additionally, if Sarah has specific questions about the design decisions made by the agents, she can request clarifications from them. The user proxy agent is designed to facilitate this feedback loop.

Also, note that even though Sarah may dismiss several options. All these engineering artifacts, from the architectures in ACEL, to simulation models and results in Simcenter Amesim, and SysMLv2 requirements validation results are available for future reference. Unlike traditional LLM chatbots that exclusively use the LLM’s knowledge, multi-agent systems make use of rigorous tools that validate the results and provide an additional layer of rigor.

All these interactions are managed by the group chat manager agent, which orchestrates the entire workflow and ensures smooth communication among the different agents. They are operating on an even higher plane of team flow than Sarah would have dreamt of. Perhaps we should describe it as agentic team flow – since AI agents are relentlessly goal orientated, never distracted and never tired.

Promise and Outlook

Multi-agent system development introduces exciting opportunities for LLMs. As LLMs better understand prompts, prompt engineering will improve, and agentic workflows will further streamline human-machine interactions, enhancing efficiency and user-friendliness.

However, due to their complexity, their development presents challenges, including increased response times and higher API costs. Emerging advancements such as smaller, faster models, reduced API costs per token, and new hardware improve inference speed. The field will undoubtedly benefit from future hardware innovations.

The potential use cases for multi-agent systems are virtually limitless, restricted only by our imagination. By adapting to changes and making decentralized decisions, multi-agent systems can improve the efficiency of complex systems like systems simulation and workflow automation in a design process.

Acknowledgments

Special thanks to Alex Graham, Daniil Chivilikhin, Daowei Yang, Dorlis Bergmann, Kefan Sun and Mike Nicolai for their contributions.