Exploring AI Agents in 2026

March 31, 2026 Development C# Azure Foundry AI Agentic AI

Exploring AI Agents in 2026

Over the past few months, I’ve been digging into building agentic AI systems in Lunch and Learns with my co-workers. You can read about that in the Omnitech Article

Things are moving fast and I’m feeling constantly behind, but that’s ok. It’s good to know about the capabilities and limitations of agents now, so we can be ready when clients start asking for AI solutions.

I admit I used AI to start this article from my notes (you can tell from the emojis), but then went through and edited it.

Why Care About Agents

But agents open the door to something more: actions, not just answers.

An interesting example is ordering a pizza through the agent’s API, or pulling data from multiple systems and actually orchestrating tasks end‑to‑end.

Microsoft is predicting more than 1 billion agents will be in used by 2028 [1]. I wonder how many will be useful, but a lot of companies are experimenting and investing.

What We’ve Explored So Far

🧱 Foundry-Based Agents

We created and invoked agents inside Microsoft Azure AI Foundry, learning how V2 agents differ, how identity works and how to call them with Microsoft’s Agent Framework.

We tried different knowledge sources including SharePoint, which needed more work and a Sharepoint instance to experiment against. I didn’t get it working.

🧪 Aspire + Microsoft Agent Framework

We experimented with Aspire and the Microsoft Agent Framework to call Foundry agents through C#. It’s very simple. However the sdks are still in preview and things change rapidly, which can break things when updating.

var aiProjectClient = new AIProjectClient(new Uri(foundryEndpoint!), new DefaultAzureCredential());
aiAgent = await aiProjectClient.GetAIAgentAsync(name: foundryAgentName);

Including the AG-UI protocol, which is a simple way to connect agents to a web UI, was helpful.

🧠 RAG Scenarios & HR Knowledge Retrieval

We uploaded the company Handbook as a knowledge source in the Foundry agent and ran simple HR Q&A tests.

It worked, but it also highlighted what better RAG solutions (like Foundry IQ) could unlock once we connect multiple knowledge sources reliably.

The Identity Problem (and Why It Matters)

One of the biggest conceptual breakthroughs for me has been understanding Agent Identity in Microsoft Entra.

Agents can run:

Attended (acting on behalf of the user)
Unattended (acting under their own RBAC identity)

This determines what your agent can access—Graph, Azure APIs, SharePoint, or internal resources.

When you publish an agent, you must reassign RBAC permissions, which was a “gotcha” that finally clicked once we looked at Entra’s JSON view.

This feels like the enterprise-ready building block that will keep future agent systems from becoming a security nightmare.

Agentic Ideas

We brainstormed some practical company automations—things that are repetitive, rules‑based, and currently cost us people-hours.

The problem was we didn’t see the value for most of them. We have people who can do these tasks and don’t mind doing them.

Many times our clients will ask for the AI agent to fix everyone. It’s necessary to ask “what are the questions/challenges you are trying to solve?” and “what if that answer won’t have the same outcome every time?”

When Should The Solution Not Be an Agent?

This was a helpful question: What should really just be deterministic code?

If a workflow:

has strict rules
has predictable inputs/outputs
requires no natural-language reasoning

…then agentic reasoning is unnecessary and going to cause issues.

The Plan Moving Forward

We want each person to build something small in their personal subscription—try everything, break things, learn what not to do—and then merge the good patterns into Omnitech Connect.

Pieces we know we’ll need:

A Foundry project
A .NET Agent project running in Azure Functions
RBAC and managed identities
A vector DB or Foundry IQ
A web UI (AG‑UI for now; DevUI is helpful for visualizing in development)
Teams integration
Documentation + architecture + repeatable patterns

We’re inching toward a repeatable recipe.

Final Thoughts

This is the least capable these LLM models and AI agents will ever be.

The pace of change is relentless, it’s time to experiment and learn so we can help our clients build real solutions that solve real problems.

Our goal is simple: be pragmatic, learn deeply, build wisely, and make sure we stay grounded in real business value—not hype.

References

🔗 AI Agents – References & Resources (2026)

Useful links I found along the way (thanks Copilot for pulling them from my OneNote)

I gladly accept BTC Lightning Network tips at strike.me/aligned

Check out my Resources Page for referrals that would help me.

Use Swan Bitcoin to onramp with low fees and automatic daily cost averaging and get $10 in BTC when you sign up.

Use the Brave browser to block ads and trackers!