Copilot agents vs Copilot Studio

Microsoft's agent strategy has multiplied entities in the last two years — Copilot agents, agents in Copilot Studio, declarative agents, autonomous agents, Copilot for Dynamics 365 — overlapping but distinct. Untangling the taxonomy is essential for picking the right tool for a specific need.

The umbrella term: agent. Microsoft uses "agent" loosely for any AI-driven assistant that handles user requests. Underneath, the implementations vary in their building blocks, hosting model, licensing, and integration surface.

Copilot Studio (formerly Power Virtual Agents). A low-code platform for building custom conversational agents. Origin: chatbots; current state: full agent builder with multi-channel publishing.

• Building blocks — topics, generative answers, actions (Power Automate, custom connectors, Dataverse).
• Channels — Teams, web, Slack, custom apps, mobile, voice via Azure Communication Services.
• Licensing — Copilot Studio messages consumed per invocation; per-tenant or per-user packs.
• Use case — replace IVR, embed in customer-facing portal, internal helpdesk, custom-domain agents.

Microsoft 365 Copilot. The end-user productivity agent embedded in M365 apps (Teams, Word, Excel, Outlook). Operates over the user's M365 graph data: emails, files, meetings, contacts.

• Building blocks — proprietary; not user-extensible at the core.
• Extension points — declarative agents and plug-ins.
• Licensing — per-user M365 Copilot licence.
• Use case — productivity in everyday M365 work.

Declarative agents. Defined in JSON manifests, deployed to M365 Copilot. A declarative agent extends M365 Copilot for a specific scenario:

• Custom instructions — system prompt tailored to a domain (e.g. "You are the HR policy expert").
• Knowledge sources — specific files, SharePoint sites, or Microsoft Graph endpoints.
• Capabilities and tools — restricted set of actions.
• Distribution — through M365 admin or AppSource.

Best for domain-specific copilots in M365 without building a full agent.

Custom engine agents (custom GPTs in M365). Beyond declarative, an agent can be built with custom logic via the Microsoft 365 Agents SDK — a C#/TypeScript SDK for building agents that run on Azure and surface in M365 Copilot.

Copilot for Dynamics 365. Pre-built agents embedded in Dynamics 365 apps:

• Sales Copilot — opportunity summaries, email drafting, meeting prep.
• Service Copilot — case summarisation, response drafting.
• Finance Copilot — variance analysis, narrative generation.
• Supply Chain Copilot — exception explanation, demand insights.

These are Microsoft-built; customers configure them but don't build them from scratch.

Autonomous agents. A 2024+ category: agents that act on triggers, run continuously, and complete multi-step tasks without conversational interaction. Built in Copilot Studio with the autonomous agent profile:

• Triggers — Dataverse events, scheduled, external API.
• Actions — tools, Power Automate flows, knowledge searches.
• Memory — short-term and long-term context.
• Approvals — humans-in-the-loop at decision points.

Use cases: inbound email triage, document review, supplier onboarding orchestration.

How they fit together.

• User asks a question in Teams chat → Microsoft 365 Copilot. If the question is HR-specific, the HR declarative agent answers.
• User is in Dynamics 365 Sales → Sales Copilot offers summaries and drafts.
• External customer chats with us via WhatsApp → Copilot Studio agent routes the conversation.
• An email arrives in support inbox at midnight → autonomous agent triages, drafts a reply, and posts for human review.

The architectural pattern: M365 Copilot is the productivity surface; Copilot Studio is the build-your-own surface; Dynamics Copilot is the in-app surface; autonomous is the event-driven backbone.

Common confusions.

• "Should I build a Copilot Studio agent or a declarative agent?" Declarative agents extend M365 Copilot's chat in M365 apps for a specific knowledge domain. Copilot Studio agents stand alone, multi-channel, often customer-facing. Different surfaces, different audiences.
• "Why are there multiple agents that overlap?" Microsoft is iterating in public. Some categories will consolidate; others will deepen. Pick based on current capability and roadmap.
• "Can I extend Microsoft 365 Copilot itself?" Yes, through plug-ins and declarative agents — but not at the core conversational engine level.

Licensing complexity. Each surface has different licensing:

• M365 Copilot — per user, ~$30/month list.
• Copilot Studio — per agent, message-pack based.
• Dynamics Copilot — bundled with the underlying D365 app licence.
• Autonomous agents — emerging metering (per task, per minute).

Cost modelling for production use is non-trivial; involve licensing specialists early.

Common pitfalls.

• Choosing the wrong surface. Building a customer chatbot in M365 Copilot when Copilot Studio is the right tool; or vice versa.
• Ignoring grounding. Agents without good knowledge sources hallucinate.
• No measurement. Agents shipped without success metrics; nobody knows if they help.
• Over-relying on a single agent type. Production AI strategy usually involves several surfaces; designing for one limits flexibility.

Practical strategy. For internal productivity, M365 Copilot + declarative agents. For external multi-channel conversations, Copilot Studio. For in-app workflow assistance, Dynamics Copilot. For event-driven multi-step automation, autonomous agents. Pieces compose; the architecture maps to who's interacting and through which surface.