Key insight
Every agent for Microsoft 365 is built one of two ways. A declarative agent is described — instructions, knowledge, and actions, defined without custom code — and runs on the shared Microsoft 365 Copilot orchestrator, inheriting its governance and compliance automatically. A custom engine agent is built — you bring your own orchestration, choose your own AI model, and can trigger workflows proactively across systems. Neither is wrong; they trade convenience and built-in governance against total control.
Before touching a specific tool, it helps to understand the fork every builder faces first: do you describe an agent and let a ready-made engine run it, or do you build the whole engine yourself? This single choice shapes everything downstream, how fast you can ship, how much governance you get for free, and how much control you keep.
1 · Declarative agents: describe, don’t program
A declarative agent provides low-code extensibility for Microsoft 365 Copilot. It lets a maker define structured workflows and actions without writing custom code: natural-language triggers mapped to predefined actions, ideal for extending Copilot with domain-specific commands. You describe the what — instructions, knowledge sources, and the actions the agent may take — and a ready-made engine, the Microsoft 365 Copilot orchestrator and its foundation models, handles the how underneath.
Because that engine is shared infrastructure Microsoft already runs and governs, a declarative agent is, in Microsoft’s own words, “governed and secure within Microsoft 365 compliance boundaries” from the moment it exists. You are not assembling security and compliance yourself; you inherit it by using the shared engine.
2 · Custom engine agents: bring your own everything
A custom engine agent takes the opposite approach: it gives developers full control over orchestration, AI models, and data integrations. Three characteristics define it. Custom orchestration — you define tailored workflows and connect to external systems yourself, rather than relying on a fixed, shared engine. Flexible AI models — you choose from foundation models, fine-tuned models, or industry-specific models to suit your exact use case, rather than using whatever model powers Copilot by default. And proactive automation — the agent can programmatically trigger workflows and take actions across enterprise applications on its own initiative, not only in response to a request.
That last point is worth sitting with. A declarative agent’s actions are triggered by natural-language requests within a defined set; a custom engine agent can decide to act unprompted, reaching into other enterprise systems as its own workflow logic dictates. That is real power, and it is also why it requires you to build the orchestration, not inherit it.
3 · Furnishing a house vs building one
An analogy makes the trade-off concrete. Building a declarative agent is like moving into a house that is already built, plumbing, wiring, and foundation all provided, and simply choosing the rooms and the furniture. You move in fast, and the building already meets code. Building a custom engine agent is like constructing the house yourself, foundations and all. You decide every wall and every pipe, which gives total control, at the cost of having to build everything by hand and be responsible for whether it meets code.
Neither approach is objectively better; they answer different questions. The furnished house answers “how fast can I have a working, compliant agent?” The self-built house answers “how much control do I need over exactly how this agent thinks and acts?”
4 · Which one should you build?
Reach for a declarative agent when your task is well described by instructions, a knowledge source, and a defined set of actions, and you want the speed and built-in governance that comes from the shared engine. Most everyday productivity scenarios, answering questions from a knowledge base, running a guided workflow, fit this description well.
Reach for a custom engine agent when you need a specific AI model the shared engine does not offer, a genuinely bespoke decision-making process, or the ability to act proactively across systems that a fixed orchestrator was never built to reach. Complex business processes, deep integrations with line-of-business systems, and scenarios requiring a particular fine-tuned or industry-specific model all point this direction.
Can your task be fully described as instructions, knowledge, and a defined list of actions? If yes, a declarative agent will likely serve you faster and with less to build. If the answer keeps coming back “no, because I need my own model or my own workflow logic,” you are looking at a custom engine agent.
5 · Glossary — every short-form term, spelled out
- Declarative agent
- An agent described through instructions, knowledge, and actions, without custom code, running on the shared Microsoft 365 Copilot orchestrator.
- Custom engine agent
- An agent built with custom orchestration, a chosen AI model, and proactive automation, giving the developer full control.
- Orchestrator
- The component that decides how an agent reasons and acts — shared and provided for declarative agents, built by the developer for custom engine agents.
- Foundation model
- A general-purpose AI model that can be used as-is or as a base for fine-tuning.
- Fine-tuned model
- A foundation model further trained on specific data to specialise it for a particular task.
- Proactive automation
- An agent programmatically triggering workflows and actions on its own initiative, not only in response to a request.
Every agent for Microsoft 365 is built one of two fundamentally different ways.
A declarative agent is described — instructions, knowledge, actions — and runs on the shared, already-governed Copilot orchestrator.
A custom engine agent brings its own orchestration, its own choice of AI model, and can act proactively across systems on its own initiative.
The trade-off is speed and built-in governance versus total control, like furnishing a house versus building one from the foundations up.
Choose declarative when instructions, knowledge, and a defined action set fully describe the task; choose custom engine when you need your own model or bespoke workflow logic.
References
- Microsoft Learn, Custom engine agents for Microsoft 365 overview. learn.microsoft.com
- Microsoft Learn, Choose the right platform (Copilot Studio) — declarative vs custom engine agent types. learn.microsoft.com
- Microsoft Learn, Customize Microsoft 365 Copilot with an agent — agent type comparison table. learn.microsoft.com