The three species
| Declarative agent | Custom engine agent | Autonomous agent | |
|---|---|---|---|
| What it is | Instructions + knowledge scope + actions ON Copilot's own engine | Your orchestrator/model behind the Copilot UI (built in Studio or code) | Agents that run on triggers/goals without a human in the chat |
| Hosting | Microsoft 365 Copilot hosts it — no infra | Yours/Studio's runtime | Studio/Foundry runtimes |
| Licensing reality | Rides the Copilot seat | No Copilot seat needed — usable from included Copilot Chat, billed via credits/consumption | Consumption + (as governance) Agent 365 |
| Typical risk | Scoped knowledge leaks (wrong site in scope) | Everything a custom app can do wrong + model behaviour | Action risk at machine speed |
| Review focus | Knowledge scope + action manifest | Endpoints, auth, data egress | Triggers, permissions, blast radius, kill switch |
The licensing fact that reshapes strategies: custom engine agents reach every employee through included Copilot Chat on consumption billing — departments can ship agent value without the per-seat debate. That same fact is why sprawl arrives faster than the seat-count suggests (next concept).
Where they come from
- Copilot Studio: both declarative and custom engine agents, low-code; post-May-2026, every new Studio agent automatically gets an Entra Agent ID (no opt-out) — the identity hook governance hangs on.
- Developer path: Teams Toolkit/VS + manifests for declarative; full code for custom engines (the Teams curriculum's app anatomy applies — an agent package is an app package with an AI section).
- Makers in the business: the Power-Platform-shaped adoption curve, with the same governance lesson: environments and DLP first or archaeology later.
What to watch (proofs)
- The census by species: the agent inventory (admin center / Agent 365 registry) classified into the three rows — if you can't produce this table, module 5's next concept is overdue.
- Licensing path per agent: which are seat-hosted vs credit-metered — the consumption report names the credit burners.
- The identity trail: new Studio agents appearing as Entra Agent IDs — creation events in the audit log; an agent WITHOUT an identity is now the anomaly.
- One manifest read per species: a declarative agent's scope block vs a custom engine's endpoint list — ten minutes that calibrates your review instincts.