Fujitsu Kozuchi AI Agent at Microsoft Ignite 2024 (EN)

Hello, this is Hiro Kobashi and Kosaku Kimura from the Artificial Intelligence Laboratory.

From November 19 to 22, at Microsoft Ignite 2024 held in Chicago, we presented a collaborative demonstration of MS’s AI orchestration tool, Semantic Kernel, and the Fujitsu Kozuchi AI Agent, much like what we did at the previous Microsoft Build 2024. We’d like to introduce it to you here.

MS Ignite 2024

MS Ignite is a customer-focused event hosted by Microsoft. Since the end of the COVID-19 pandemic, it has been held annually around November. This year’s event saw approximately 14,000 attendees onsite—roughly triple the number from last year—and more than 200,000 online participants, making it a huge event. Over 80 product updates were announced among more than 800 sessions, and although there were even days when snow accumulated in a chilly Chicago, every session room was brimming with enthusiasm.

The theme Agentic World at MS Ignite 2024 left a strong impression. Many sessions were abuzz with product updates and use cases aimed at realizing AI agents and multi-agent systems. In the keynote by Satya Nadella, the CEO, it was declared that Copilot is the UI for AI. Copilot stands closest to the user experience. Moreover, behind the scenes, MS showcased their product stack at this event timing, including Copilot Studio—an environment for developing autonomous agents—and Azure AI Foundry—an execution environment. Their ability to unveil these at once highlights Microsoft’s strength in this area.

Summarizing our impressions about agents:

• Everything is being called an “agent”: Personally, I’m of the view that if a specialized generative AI is sufficiently usable on its own, calling it an agent is not a problem. However, since many specialized generative AIs are not fully usable as-is, they likely need to be integrated with tools and planning as described by the concept of Agentic AI.
• Multi-agent systems are becoming a reality: For example, Toyota’s use case describing their practice of building multiple agents in Azure—each possessing specialized knowledge of particular powertrain technologies—was fascinating. Similarly, Cognizant’s use case showcased multiple agents connected via a Directed Acyclic Graph (DAG) interacting in real scenarios, which I found highly intriguing. On the other hand, complex, multi-directional interactions among agents seem still relatively rare.
• Autonomy here often means “acting on pre-defined triggers,” not just “acting of one’s own accord”: Many cases gave the impression that what they call autonomy is akin to a stored procedure triggered by preset conditions, rather than self-motivated action.

Breakout Session: Productive AI with Semantic Kernel

As with MS Build, this time we also held a breakout session with the Semantic Kernel team.

This time, instead of focusing on Fujitsu Composite AI as before, we showcased its evolution into Fujitsu Kozuchi AI Agent.

Since our session was on the last day of the event, we didn’t have a completely full house, but we were happy to reach a tech-savvy, highly focused audience.

Here’s a brief introduction of our presentation:

We believe that in the near future, the way humans interact with AI will change. Specifically, rather than the current “ask and answer” (passive) model that includes generative AI, we’re moving toward a model where AI “thinks and acts on its own,” blending naturally into human-to-human interactions and providing new insights and flashes of inspiration (proactive). As a result, AI enters conversations, acts cooperatively, and suggests new knowledge and ideas—thus enabling creative collaboration between people and AI. We call this shift “Beyond Chat.”

Based on this concept, we developed the Fujitsu Kozuchi AI Agent. Here’s a YouTube video showing an example of how it functions as a meeting agent.

www.youtube.com

The selling point of the Fujitsu Kozuchi AI Agent is that it has already realized the concept of “thinking and acting on its own.” During meetings, the agent listens to the conversation continuously, identifies tasks that seem to need solving, and automatically executes them to present the results.

This is the system flow diagram. In simple terms, it consists of three parts:
“①–②” listens to the conversation in the meeting,
“③–⑦” is where the Agentic AI creates tasks and executes them to obtain results (the part previously handled by Fujitsu Composite AI),
“⑧–⑨” returns the results.
The key point is that “①–②” and “③–⑨” are fully asynchronous. This means it’s not a simple “ask and answer” system like before. Of course, if you mention the agent directly in the chat, there’s also a path to get answers in the traditional Q&A manner.

Check out the Microsoft Tech Blog’s case study for more details if you’re interested!

Customer Case Study: Fujitsu Kozuchi AI Agent Powered by Semantic Kernel @ MS Dev Blog

Also, during Ignite, we had the chance to meet with Matthew, who was previously the PM for Semantic Kernel. Matthew is now the PM for Azure AI Agent (we share a similar interest!). He praised our Agentic AI approach: “Listening to conversations, using tools, and providing responses is amazing!” He was impressed again by how we developed this in just a few months—just as we did at MS Build.

At the end, we took a commemorative photo with all the presenters :)

Fujitsu Will Continue to Focus on AI Agents

As a side note, in Matthew’s presentation about Azure AI Agent Service, Fujitsu was featured prominently as a customer. This is not only because we continue to collaborate with the Semantic Kernel team, but also because another Fujitsu team—separate from ours—participated in the Azure AI Agent Service Co-Building Program.

Fujitsu will continue to invest heavily in AI agents, so please stay tuned!