- Subject
- Microsoft 365 & SharePoint
- Filed
- Source
- Koskila
- Loans
- 926
Microsoft 365 Copilot and Foundry: Key Design Choices for Building Agents
A session at aMP Day Rennes 2026 examined when Microsoft Foundry, Copilot Studio and code first agents fit into a Microsoft 365 AI architecture.

The design question for Microsoft 365 AI projects is shifting from “How do we write a better prompt?” to “Where should the system’s intelligence and orchestration live?” That was the focus of a session at aMP Day Rennes 2026: Microsoft 365 Copilot can serve as the user interface, while Microsoft Foundry provides tools for building and coordinating more capable agents.
The distinction matters for teams deciding whether to stay with low code tooling or invest in a software engineering approach. A simple agent can amount to little more than a prompt wrapped in an interface. More demanding systems may need deliberate model selection, evaluation, orchestration, memory and connections between agents. Those choices bring added complexity, so the architecture should match the work the agent needs to perform.
Copilot as the interface, Foundry as the engineering layer
Microsoft 365 Copilot is becoming an interface for AI experiences across Microsoft 365. Microsoft Foundry gives developers a place to select and evaluate models, orchestrate agents, combine tools and build systems that go beyond a conventional chatbot. The Microsoft 365 Agents SDK can connect those systems with Microsoft 365, Teams and Copilot.
That arrangement points to a division of responsibilities: people can interact through familiar Microsoft 365 surfaces, while developers build the underlying agent capabilities with the tools and patterns the use case requires. The session’s framing was that Copilot may be the interface, while Foundry increasingly supplies the “brains” behind some experiences. That is a direction to watch, rather than a reason to put every AI task on a code first platform.
For administrators and technical leads, the practical decision is where a project’s requirements outgrow a simpler builder. The session highlighted Copilot Studio as a sensible option when low code capabilities meet the need. A code first design can become more useful when the team needs finer control over models, evaluation, orchestration or memory. Every added component should solve a real requirement; complexity without a clear purpose can make an agent harder to operate and maintain.
What the session set out to demonstrate
Rather than focus on prompt engineering, the presentation treated agent development as software engineering. It proposed building an agent from the ground up and using that work to examine several parts of Microsoft’s AI stack:
- Choosing models and evaluating their suitability.
- Orchestrating agents and tools to carry out work.
- Considering agent memory and Work IQ.
- Connecting custom engine agents with Microsoft 365.
- Exploring collaboration between agents.
- Weighing the tradeoffs between low code and code first approaches.
The intended emphasis was practical: code, demonstrations and architectural tradeoffs, including choices made for learning that might not belong in a production design. That distinction is useful for teams reviewing conference examples. A demo can make a pattern easy to understand, but implementation decisions still need to be judged against maintainability, scale and the actual requirements of a deployed system.
Architecture questions for Microsoft 365 teams
The session’s takeaways centered on knowing when Foundry adds value. Model choice, evaluation, orchestration and memory can be important when they address a concrete need. They can also become unnecessary machinery if a simpler agent already does the job. Teams should be able to explain what each part of the architecture contributes and what operational burden it introduces.
The same test applies when deciding whether Copilot Studio remains productive. Low code tools can make it easier to assemble an experience, but a project may eventually need more control than they offer. At that point, a code first approach may pay off. The useful question is not whether one tool is universally better; it is where the project’s requirements make the next layer of engineering worthwhile.
Scalability and maintainability deserve attention before an agent moves beyond a demonstration. A pattern that looks impressive on stage may create avoidable work when teams need to support it over time. The session aimed to help attendees spot those tradeoffs, avoid costly architectural missteps and leave with techniques they could try with Foundry, custom engine agents, Work IQ and Microsoft 365 extensibility.
From conference sessions to applied work
The Rennes presentation was part of a busy stretch for the speaker, who said they had presented at four events over five to six weeks. Those events included the HELish Summit, AI Maitri Global Summit 2026, aMP Day Rennes 2026 and the AI-Native Workplace Summit of 2026. The speaker described the preparation as demanding, but rewarding, and credited employers Valo, Omnia and Meltlake with supporting conference participation during different stages of their career.
The route to speaking began with attending technology community events, including SharePoint Saturdays in the United States and Montréal. Those gatherings offered a chance to learn from other people’s technical and personal experiences, and eventually encouraged the speaker to submit sessions of their own. For attendees, that community exchange is part of the value: real project stories can show where a technology helps, where it adds friction and which tradeoffs deserve attention.
The presentation slides are available for download. For Microsoft 365 teams planning agent work, the central takeaway is to start with the required capability, then choose the lightest architecture that can deliver and support it. As Copilot, Foundry and agent development continue to evolve, that boundary between a useful low code build and a system that needs software engineering will be worth reassessing.