The Journey to Multi-Agent Innovation

At ServiceNow, we embarked on an ambitious journey to redefine how AI systems collaborate in enterprise environments. Through our partnership with Microsoft, we set out to create something beyond traditional integration – a true multi-agent system across platform that could work effectively alongside human teams. While NowAssist and Microsoft Copilot were already capable of basic task handoffs, we envisioned a future where AI agents could work together as seamlessly as human colleagues, sharing context and coordinating complex activities in real-time.

Tackling the P1 Incident Challenge

Late 2024, a small team devoted intensive effort to understanding the intricacies of multi-agent coordination. One of our greatest challenges was developing a system that could maintain context and coherence across different AI platforms while ensuring reliable, accurate outcomes. We needed more than just a technical integration – we required a solution that could support the full lifecycle of a real-world challenge.

We chose to focus our initial implementation on one of the scenarios in enterprise IT with potential tremendous value: P1 incident management.

These critical situations traditionally require hours of intense collaboration, often conducted through rapid-fire verbal communications that leave behind incomplete documentation. Through careful analysis, we discovered that Copilot’s advanced transcription capabilities, combined with NowAssist’s deep integration with ServiceNow systems, coordinated by Semantic Kernal’s orchestration capability could transform this process. The challenge wasn’t just about recording what happened – it was about creating an intelligent system that could understand, act upon, and learn from these high-stakes situations.

Tackling P1 Incidents Through Multi-Agent Architecture

We chose to focus our initial implementation on one of the high stakes and high value scenarios in enterprise IT: P1 incident management. These critical situations traditionally require hours of intense collaboration, often conducted through rapid-fire verbal communications that leave behind incomplete documentation. Through careful analysis, we discovered that combining Copilot’s advanced transcription capabilities with Now Assist’s deep integration into ServiceNow systems could transform this process.

The foundation of our multi-agent framework lies in the manager agent architecture, which serves as the orchestrating intelligence of our multi-agent system. It maintains a comprehensive list of actions to be executed, understands the capabilities of each sub-agent at its disposal, and manages the overall state of incident response activities. Our implementation leverages Semantic Kernel’s orchestration capabilities, allowing these different components to work together seamlessly.

When a critical incident occurs, the system automatically generates a Microsoft Teams bridge call through the ServiceNow platform, gathering the major incident management team in a unified space. During these live P1 calls, this manager agent springs into action. Copilot acts as an intelligent observer, capturing and interpreting verbal communications in real-time and highlighting the actions. This information flows to Now Assist and feeds into the manager agent, which processes it to trigger appropriate actions within ServiceNow platform.

What makes this approach particularly powerful is its flexibility in handling data queries and escalations. Rather than constraining the system with rigid workflows, we developed an adaptive framework where Now Assist can autonomously assess situations and take appropriate action. For instance, when team members discuss potential scale of impact, Now Assist can independently query ServiceNow instances to gather relevant data. If the analysis reveals significant issues, it can initiate escalation procedures with humans in the loop.

This dynamic approach, guided by the manager agent’s orchestration, allows the system to adapt to each incident’s unique characteristics while maintaining consistent accuracy and reliability. The manager agent continuously evaluates the situation, selecting and coordinating the most appropriate sub-agents for each task. This orchestration ensures that all actions are not only executed efficiently but also properly documented and tracked throughout the incident lifecycle.

Potential Business Impact

Our proof of concept demonstrated the remarkable potential of collaborative AI in incident management. During controlled testing sessions, we observed how the synergy between Now Assist and Copilot could transform the incident resolution process. The system successfully maintained contextual awareness across platforms, ensuring that communications in Teams were seamlessly reflected in ServiceNow and vice versa. This synchronization addresses one of the most persistent challenges in incident management: the problem of fragmented information and inconsistent updates across different platforms.

The proof of concept revealed promising results in addressing the long-standing challenge of post-incident documentation. The collaborative interaction between our AI agents demonstrated the capability to automatically generate comprehensive incident reports and knowledge base articles. This automation could potentially transform what has traditionally been a time-consuming manual process into a streamlined, reliable system for knowledge preservation and transfer.

The system’s ability to maintain consistent context across platforms while facilitating real-time collaboration between AI agents suggests substantial improvements in incident resolution efficiency. Furthermore, the automated documentation capabilities could lead to more comprehensive and accessible knowledge bases, enabling faster resolution of similar incidents in the future.

Future Directions and Process Framework

Looking ahead, we see tremendous potential in expanding this collaborative model across platforms and even across enterprises. The introduction of new AI capabilities and enhanced reasoning models promises even more sophisticated interactions between agents. We’re particularly excited about exploring how emerging technologies can enhance the system’s ability to learn from past incidents and proactively suggest preventive measures.

Our experience has taught us that successful multi-agent implementation requires careful attention to both technical integration and human factors. There is still tremendous work ahead to further expand the agentic capability and continue building sophisticated cognitive architecture with strong guardrails. With the prominent shift to Agentic AI, the work interface will have to change as well – we need to re-think how future human/AI collaboration should look like.

As we continue to develop this system, we remain focused on maintaining the delicate balance between autonomous operation and human oversight, ensuring that our AI collaboration enhances rather than complicates the incident management process.

Conclusion: A New Chapter in Enterprise AI Agent

Our proof-of-concept implementation has demonstrated the significant potential of AI collaboration in enterprise environments. Through the integration of Now Assist and Microsoft Copilot, we’ve shown how a multi-agent system can enhance incident management while building stronger foundations for knowledge capture and transfer. While still in early stages, this implementation suggests that with thoughtful architecture and design, AI agents can become valuable partners in enterprise teams, working alongside humans to improve operational outcomes.