Separate behavior from platform
Define the system's responsibilities before choosing where it runs. The creator compares architecture options during discovery. You confirm the target architecture before it prepares the complete package.
You can add an agent without building a new web application. Choose where custom agent code runs separately from how you write it. And check whether an existing backend can support the interface you need.
Start with what you already have
Check whether existing applications and platform capabilities can support the first journey. Reuse them when they meet the functional and operating requirements. A new repository can still depend on services the enterprise already runs.
Microsoft's technology planning guidance starts by checking whether a ready-to-use agent meets the requirements before comparing build options. During discovery, the creator checks for reuse first and identifies which confirmed requirements need custom behavior.
Execution options the creator compares
The creator must verify that the selected services work together before delivery.
| Path | When to evaluate it | What your team must still define |
|---|---|---|
| Existing or prebuilt capability | It may already meet the user journey. | Fit, access boundaries, ownership, and any gaps. |
| Copilot Studio | A low-code agent or workflow could meet the integration and channel requirements. | Connections and permissions, business approvals, environment lifecycle, and evaluation. |
| Foundry prompt agent | Instructions, model configuration, and permitted tools could express the behavior. | User access, knowledge permissions, tool policy, and acceptance criteria. |
| Foundry hosted agent | Custom code is needed and managed execution could fit. | Code and protocol choices, business state, approval semantics, and operational requirements. |
| Application-owned execution | Existing application integration or runtime control is a confirmed requirement. | Hosting, authorization, durable state, deployment, and telemetry. |
| Explicit workflow or bounded inference | Steps are known or one model response meets the task. | Output validation, failure handling, and integration with the existing process. |
Copilot Studio provides low-code agent and workflow authoring. Foundry Agent Service offers configured prompt agents and managed hosting for custom agent code. Use their current documentation to verify the exact feature and connection being considered.
If custom code remains necessary, compare managed execution with application-owned execution. The skill includes recipes for Agent Framework and the GitHub Copilot SDK; choosing either framework does not determine where the code runs.
Choose the channel independently
Identify where the user or caller already works. It may be an existing application, a supported platform channel, or a background process. Build a custom client only when a confirmed requirement calls for one.
Verify the chosen channel's audience and authentication model. Then decide whether it provides the needed experience or requires a client you must own. Choose the client separately from the agent runtime.
For example, a maintenance investigation might live inside the existing service application. An event-driven document workflow might need no interactive client. Test either choice against your requirements.
Keep knowledge and actions distinct
A knowledge source supplies evidence. A business tool performs an operation. Even when both reach the same system, they need separately defined permissions and responsibilities.
When several sources are needed, record what each contributes and how the system enforces access. Select the model and inference boundary separately from the knowledge layer and agent runtime. Reuse suitable existing deployments rather than assuming the project needs a new one.
Verify the proposed connections
Before you confirm an architecture, the creator checks product facts that affect the design against current Microsoft Learn documentation and the relevant specialist guidance. The package records unresolved checks and their owners.
Verify the selected channel, protocol, identity, networking, and lifecycle. Account for licensing and region availability where they affect the design. Distinguish preview capabilities from generally available ones. Verify any adapter the integration needs.
If a connection remains unverified, delivery work that depends on it must wait. Architecture approval does not turn an unverified integration into a supported one.
Give discovery your constraints
Give the creator requirements and existing standards:
/ai-agent-creator Design [capability] inside our existing service application. Reuse its
identity and business APIs. Compare Copilot Studio, managed Foundry execution,
and application-owned execution where they fit. For every new UI, API, or host,
name the requirement that makes it necessary.
Set up the workspace to begin discovery. If there is an application already in place, use the existing-application guide to prepare its context.