Declarative agents or custom engine agents: choosing how to extend Copilot

Not every extension needs an agent. Microsoft describes agents as specialized assistants that add knowledge and actions to Copilot¹, but two smaller pieces often come first. A Copilot connector brings content from another system into Microsoft Graph so Copilot can find it, and an action lets Copilot call an API to read or change a record. If employees only need Copilot Chat to find tickets from your service desk, a connector on its own may be enough.

You need an agent when a defined audience needs a dedicated experience: consistent instructions, curated knowledge, a set of permitted actions and conversation starters that explain what the agent is for⁴. At that point the question becomes which kind of agent, and that turns on who runs the orchestration and the model.

With a declarative agent you write configuration, not an engine: the agent's purpose and instructions, the knowledge sources it may ground in, the actions it may call and the conversation starters users see. Copilot plans each turn, retrieves content within the user's permissions, decides when to call an action and composes the reply.

That division of labor is the main advantage. Because the agent runs on Copilot's infrastructure, model and orchestrator, it adheres to the same security, compliance and responsible-AI requirements as the rest of Copilot¹. The trade-off is control: you cannot swap the model, add your own planning logic or have the agent open a conversation by itself. When a declarative agent misbehaves, your levers are the instructions, the scope of its knowledge and the design of its actions.

Microsoft's guidance gives a short list of reasons¹: business rules too precise to leave to a general orchestrator, a need for your own or a domain-specific model, an assistant that serves a whole Teams channel or meeting, an existing bot you want to bring into Copilot, availability outside Microsoft apps, and proactive messages triggered by events rather than prompts.

Every one of those reasons brings work with it. You host and secure the orchestrator, manage model versions, implement content filtering and logging, and answer your security team's questions about where prompts and documents travel. Copilot Studio can host a custom engine agent for you, which removes the infrastructure burden but not the design responsibility.