AI
What Agency Actually Means in Code
Seth Juarez pulled up a dictionary before touching a slide. “Agency is the ability to take action or to choose what action to take.” That’s it. The rest is implementation.
An LLM takes tokens in and outputs one token at a probability distribution. Loop that, and you get text. Give it a system prompt that says “classify this as support or sales,” and the token it picks can branch your code. Give it a function schema and it outputs an object shaped exactly like a function call, telling your program what to run next. That is the full mechanism.
“We effectively have a new programming control structure.” — Seth Juarez
The Agent Anatomy: Instructions, Triggers, Tools, Knowledge
Kendra Springer mapped Seth’s code walkthrough onto the product layer. Every agent has five bones: instructions (the system prompt), triggers (what invokes it), tools (what it can call), knowledge (the data it reads from), and an orchestrator that decides which combination to use per query.
The mapping holds across both products. In M365 Copilot’s Embedded Agent Builder, the orchestrator and model merge into M365’s proprietary LLM service. In a Copilot Studio custom agent, you get a rule-based or generative orchestrator and access to 11,000 models in Azure AI Foundry. Same pattern, different controls exposed.
Choosing Between Copilot Studio and Azure AI Foundry
The question customers ask most: “Which tool should I use?” Seth Juarez’s answer from Azure AI Foundry: start at Copilot Studio, see how far it gets you, then drop to Foundry only where it breaks. Foundry is for agents that need to live inside an existing dev workflow, a brownfield app, or a specific model. Copilot Studio covers everything else, including event-driven autonomous agents, MCPs, and custom RAG over Azure AI Foundry indexes.
Both platforms run the same underlying pattern: instructions, tools, execution flow. The ceiling difference is how much configuration responsibility the builder accepts.
Copilot Studio Demos: Procurement and Autonomous Supply Chain
Kendra Springer ran the procurement advisor built in 90 seconds (32:35) live: a natural-language prompt, a SharePoint knowledge source, and a working agent that checked statements of work against the company rate card. The second build was autonomous. The autonomous supplier discovery agent (37:29) triggered on a simulated inventory surge, pulled ERP records, searched the web for additional suppliers, and pushed a ranked proposal report to Teams.
Seth Juarez built the Azure AI Foundry live agent build (44:48) in under two minutes with GPT-4o and Bing grounding, then showed the code-level event handler and the evaluator library for automated agent tests.
Multi-Agent Orchestration and the EY Case Study
The Contoso banking multi-agent handoff demo (50:25) showed a root agent routing a fund-transfer request to a dedicated transfer sub-agent, then picking up a mortgage agent mid-session for a payment-due query, and repeating both flows in Spanish without re-configuring anything. Adding a new sub-agent required one click and no republish.
EY’s Mark Luquire put real numbers behind the capability. EY’s Portal One indexes more than 21 million legal and tax documents across 150 countries. It was built on Copilot Studio and Azure AI Foundry, handles queries across 12,000-plus US tax jurisdictions, and shows citations alongside every answer so tax professionals can verify before acting.
Notable Quotes
We effectively have a new programming control structure. Seth Juarez · ▶ 14:13
11,000, I saw today. That’s insane. Seth Juarez · ▶ 19:01
Not simple, highly powerful. Kendra Springer · ▶ 25:16
Key Takeaways
- An agent is an LLM that returns function calls instead of text, giving code a new kind of conditional branch.
- Copilot Studio and Azure AI Foundry both implement the same model-instructions-tools pattern at different abstraction levels.
- EY’s Portal One indexes 21 million legal and tax documents, proving retrieval agents scale far beyond ‘simple.’