- Subject
- Microsoft 365 & SharePoint
- Filed
- Source
- Koskila
- Loans
- 494
AI Maitri Global Summit 2026 Session Examines Microsoft’s Agent Building Stack
A manufacturing focused demo explored how Microsoft Foundry, Copilot Studio and Work IQ can support agents that act on business workflows.

Building an AI agent for operational work takes more than connecting a chatbot to a language model. That was the engineering question behind a manufacturing focused session at AI Maitri Global Summit 2026: how can Microsoft’s tools help teams move from conversation toward systems that support processes and decisions?
The session, titled “M365 Copilot is the UI, Foundry is the Brain, but what turns intelligence into action?”, used a working agent as a way to examine the Microsoft AI stack. Its focus was on practical architecture: choosing models, evaluating their performance, coordinating agents and tools, managing memory, and connecting the resulting system to Microsoft 365 experiences.
From a demo to an engineering discussion
The presentation was submitted to the summit’s Manufacturing track. Its premise was that industrial organizations need more than a model that can answer questions. They need systems that can use relevant context and reasoning to support operational workflows, automate parts of a process and help people make informed decisions.
That distinction shaped the session’s approach. Rather than treating prompt design as the main engineering task, it framed agent development as software engineering. Model selection, evaluations, orchestration and memory all affect how an agent behaves, how it fits into a larger system, and whether the added complexity is justified by the work it performs.
Microsoft 365 Copilot was presented as an interface for AI experiences across Microsoft 365. Microsoft Foundry, meanwhile, was described as the environment where developers can select and evaluate models, orchestrate agents, combine tools and build systems beyond conventional chatbot interactions. The Microsoft 365 Agents SDK provides a connection back to Microsoft 365, Teams and Copilot.
Those roles matter when teams decide where to build. Copilot Studio may be sufficient for some agent scenarios, while a code first approach may offer more room when requirements call for custom orchestration or deeper control. The session aimed to help attendees recognize that boundary and weigh the maintenance and architectural tradeoffs before a demo becomes a production design.
What the session explored
The demonstration covered several parts of the stack and how they can work together:
Model choice and evaluation: Foundry can support decisions about which model to use and how to assess it. The point is to connect those choices to the intended business outcome, rather than adding evaluation work without a clear reason.
Orchestration and memory: The session examined patterns for coordinating agents and tools, along with ways to use memory. These choices influence how an agent handles more involved tasks and how maintainable the system is.
Copilot Studio and code: Attendees were invited to consider when low code tooling remains productive and when writing code begins to pay off. The right choice depends on what the agent must do and how much control its builders need.
Work IQ and custom engine agents: The session explored how these capabilities could work with Foundry to provide context and support action oriented experiences.
Agent to agent collaboration: The demo also looked at collaboration between agents as one possible orchestration pattern, alongside the practical question of whether a pattern will scale and remain maintainable.
For administrators and development teams, the useful test is whether each layer solves a real requirement. A richer architecture may offer more control, but each component also adds design and maintenance work. The session’s emphasis was on understanding where those tradeoffs improve an outcome and where they amount to complexity that the use case doesn’t need.
A last minute build and an unusual venue
The slides and demo were completed only minutes before the session. The team had run into problems involving sub agent orchestration, its use of Work IQ, and the way a Copilot Studio agent accessed an API. The account offers a reminder for teams preparing live demos: an agent’s behavior depends on how its parts connect, so integration details can matter as much as the model or prompt.
The online format made it possible to present from a childhood room during a weekend visit to the presenter’s parents. The room had since become a storage space, adding an unexpectedly personal setting to a technical session. The presenter also reported an engaged audience, with questions and feedback after the demo.

Slides and next steps for teams
The session materials are available as a PDF: download the slides. Teams reviewing the ideas can use the deck to compare their own agent requirements with the capabilities discussed: model evaluation, orchestration, memory, Work IQ, custom engine agents and connections to Microsoft 365.
The practical takeaway is to start with the work an agent must support, then choose the tools and architecture that fit. Copilot can provide the user facing experience, while Foundry offers options for building the intelligence and coordination behind it. The engineering decision is knowing which capabilities the workflow needs, and which ones the team can maintain.