Welcome to the fifth episode. In this chapter, we’ll share practical tips and best practices to improve agent behavior: how to iterate on instructions and prompts, how to understand the agent’s reasoning, and how to extend its capabilities by adding more tools or additional MCP Servers.
Let’s look at a few tips, and at the end, how everything comes together:
🌱 PROMPT COACH, CHATGPT, AND COPILOT FOR PROMPTING
First and foremost, it’s absolutely essential to clearly define the instructions and the prompt when we create an agent and interact with the MCP Server.
That said, it doesn’t mean we have to be world-class experts in configuring roles and agents right away. You can always rely on the same tools we already have available and are using today. They’re an excellent starting point to understand how to interact and, more importantly, how the tool actually works. They also help reduce back-and-forth, improve clarity, and catch ambiguities or conflicting instructions early on.
You can ask for help, advice, or examples, or request a review to see whether the wording is clear and correct. And if you notice it’s too much, or not enough, you can always go back and edit it.
In the end, the key to configuring an agent is to test and refine — and keep testing and refining — until it responds the way we expect.
🌱 TALK TO YOUR AGENT
One of the things I struggled with the most at the beginning was understanding how the model assigned to the agent actually reasons. And here’s a tip that helps a lot: if the outcome isn’t what you expected, ask it why it took that path.
Instead of assuming it “did it wrong,” ask it to explain its logic: why it chose that tool or those steps, what information was missing, what it assumed (and whether that assumption is correct), and what alternative it would have tried.
That conversation allows you to fine-tune the instructions and the prompt, and most importantly, understand what the agent needs in order to behave the way you expect.
🌱 ADD MORE TOOLS (OR MORE MCPs) TO YOUR AGENT TO IMPROVE THE EXPERIENCE
Another interesting option is to expand the agent’s capabilities by connecting it to more tools, or even to more than one MCP.
In many scenarios, it’s not enough for the agent to interact with just a single system. We may need it to query different data sources, navigate across multiple applications, or handle different types of tasks from a single entry point.
In those cases, you can add more tools or connect additional MCPs to the same agent, so it can orchestrate multiple interactions within a single conversation.
This allows, for example, a single agent to review information in FnO, consult documentation in SharePoint, and then execute an action in another system, without the user having to switch context.
The trick, as always, is to clearly define the agent’s role and boundaries, so it knows when and how to use each tool. And when interacting with the agent, stating what we need in a concise way and using a well-crafted prompt will always help.
Prompt:
Check if there is an open invoice with due date in 2017 for Sunset in USMF and send an email to the collections contact asking what time I can go to the office for a meeting. Write the email first in English and then in Spanish in the same email.
BONUS TRACK:
🌱 EXTENDING THE FnO SIDECAR COPILOT WITH MCP
What’s really useful is extending the Copilot that already comes embedded in D365 FnO (the sidecar).
From Copilot Studio, we can add additional capabilities and connect it to the Dynamics 365 ERP MCP, so that our Copilot inside the system doesn’t just answer questions, but can actually interact with FnO using all the available tools and context — all from a single place.
This opens the door to scenarios where the user continues working inside the ERP, using the embedded Copilot, but with much more powerful behavior: form navigation, action execution, and task orchestration, all within the same experience.
Below I’m sharing a video showing how to configure it, but let me walk you through a few key points first:
1. From the list of agents, we’ll find our FnO Copilot named “Copilot for finance and operations apps”:
2. Inside the agent, we’ll find the “Configuration” section:
In the “Generative AI” section, make sure that during orchestration the responses can be dynamic, using the available tools and knowledge as needed:
3. Revisar que el MCP de FnO esté conectado. Sino, al momento de interactuar desde FnO nos aparecerá el mensaje con el link a la página:
We can see this with an agent that helps us with tasks related to the employee workspace, such as entering timesheets, requesting time off, or submitting travel expenses.
I hope this series left you with ideas and, above all, the urge to try things out. Because even though the series is ending, this is where the best part begins: building your own agent ✨ See you in the next episodes, there’s sooo much more to come 😉🫶