Feature Comparison
Detailed side-by-side comparisons of Microsoft AI technologies. Use this page after you’ve clarified requirements in the Decision Framework, so the tables confirm trade-offs (rather than driving the decision by brand names alone).
Sanity check before you compare: If the work is deterministic, repeatable, or can be expressed as a clear function/workflow, don’t use an agent. Use a workflow, API integration, or traditional code. Agents are best when the task is ambiguous, multi-step, and conversational, where the steps can’t be fully specified up front.1
Table of contents
- Comprehensive Platform Comparison
- Developer Agent Comparison
- Workflow Orchestration Platform Comparison
- When to Use Each Workflow Technology
- Agent Governance Layering Guide
- IQ Layer Comparison
- Lifecycle & Operational Readiness Comparison
- Data Grounding Technology Comparison
- Local and Edge AI Comparison
- Agent Development Approach Comparison
- Custom Engine Agent Tool Comparison
Comprehensive Platform Comparison
Quick view of core platforms (M365 Copilot, Studio, Foundry, Agent Framework, Logic Apps) across UX, governance, orchestration, and hosting trade-offs.
How to use this section: Start with the columns that match your trust boundary and channel needs, then compare orchestration and governance features to validate whether you should stay in SaaS, step into low-code, or move to pro-code.
| Feature | M365 Copilot | Copilot Studio | Microsoft Foundry (Azure) | Microsoft Agent Framework | AG-UI Protocol (Preview) | Microsoft Foundry Agent Service | M365 Agents SDK | Azure Logic Apps | |
|---|---|---|---|---|---|---|---|---|---|
| User Experience | M365 apps (Frontier creation agents Preview2) + mobile parity for custom engine/message-extension agents3 | Custom channels | Custom apps | Embedded in apps | Protocol-based web/mobile UI | Custom apps | Copilot/Teams/Web | Custom/Conversational | |
| Build Approach | No-build (consume) | Low-code to pro-code | Code-first | Pro-code (orchestration SDK) | Pro-code protocol integration (ASP.NET Core, FastAPI) | Code-first | Pro-code | Visual designer + code | |
| Data Boundary | M365 tenant | M365 or Azure | Azure | Any | Inherits host runtime | Azure | M365 or Azure | Azure | |
| Governance | Tenant-integrated (automatic) + registry lifecycle (publish/activate/deploy/pin/block/remove/delete/owner transfer/export)4 | Tenant or workload | Workload-tailored (custom) | Application-level | Inherits host app controls; approvals via middleware | Workload-tailored (custom) | Tenant or workload | Azure RBAC | |
| Admin Center | M365 admin center | Power Platform admin | Azure RBAC | Application-level | Host platform (Azure/App) | Azure RBAC | M365 admin center | Azure portal | |
| Licensing Model | Per-user (/month) | Metered messages | Azure consumption | Open-source (free) | Open-source adapters (no license) | Azure consumption | Included in M365 | Consumption or Standard | 5 |
| Extensibility | Via Studio/SDK + Copilot Search API (Preview, OneDrive hybrid search)6 | Plugins, connectors | Full custom | Workflow orchestration | Seven protocol features (streaming, backend tool rendering, human approvals, generative UI, shared/predictive state) | Full custom | Full custom | 1,400+ connectors | |
| Deployment | Microsoft-managed | Microsoft-managed | Self-managed | Self-managed (SDK) | Self-hosted endpoints (ASP.NET Core, FastAPI) | Microsoft-managed | Self-managed | Self-managed (Azure) | |
| Time to Value | Immediate | Days to weeks | Weeks to months | Weeks (with dev skills) | Weeks (requires pro dev + UI build) | Weeks to months | Weeks | Days to weeks | |
| Skill Level | End user | Maker to developer | Developer/engineer | Developer (C#/Python) | Developers (front-end + back-end) | Developer/engineer | Developer | Developer | |
| Orchestration | Built-in | Built-in | Custom | Workflow-based (Executor/Edge) | Inherits Agent Framework orchestration | Managed orchestration | BYO orchestrator | Visual workflow | |
| Checkpointing | N/A | No | Custom | Yes (built-in) | Inherits from orchestrator | No | Via orchestrator | State management | |
| AI Agent Workflows | N/A | Via agent flows | Via Microsoft Foundry Agent Service7 | N/A | N/A (UI protocol) | ✅ Yes | N/A | ✅ Yes (Preview)8 | |
| MCP Server | N/A | ✅ Yes (Preview) | ✅ Yes (MCP tool)9 | Consumer (GA) | N/A | ⚠️ Custom | N/A | ✅ Yes (Preview)10 | |
| Optimized For | Productivity at scale | Speed to market + connectors | Latency & control | Workflow orchestration | Bespoke agent UI with streaming & approvals | Managed agents | Pro-code flexibility | Enterprise integration | |
| Latency Profile | <1s (M365 apps) | <1s (managed platform) | <100ms (direct API) | Workflow processing | SSE streaming (depends on backend) | <100ms (direct API) | <1s (M365 integration) | Workflow processing | |
| Infrastructure Model | Microsoft-managed | SaaS (managed) | PaaS (self-managed) | SDK (self-managed) | SDK bridging host + client | PaaS (self-managed) | SDK (self-managed) | PaaS (self-managed) | |
| Best For | Broad productivity | Custom agents | Custom AI apps | Workflow orchestration | Custom branded experiences and generative UI atop Agent Framework | Managed agents | Pro-code extensions | Enterprise integration + AI |
Developer Agent Comparison
The comparison below covers four developer-agent roles: asynchronous repository work, parallel development-session coordination, Azure reliability operations, and supported application-stack modernization. GitHub Copilot cloud agent is the remote worker, GitHub Copilot app is the dispatch desk, Azure SRE Agent is the reliability engineer, and GitHub Copilot Modernization is the renovation crew. Their capabilities overlap; their primary jobs differ.
| Feature | GitHub Copilot cloud agent | GitHub Copilot app | Azure SRE Agent | GitHub Copilot Modernization |
|---|---|---|---|---|
| Primary Role | Asynchronous repository research, planning, code changes, and pull requests in a GitHub Actions-powered environment | Desktop control surface for parallel agent-driven development and GitHub pull-request lifecycle management | Azure operational investigation, diagnosis, and approved remediation | Modernization solution delivered through complementary IDE extensions and the Modernize CLI—not a standalone agent product |
| Execution Surface | GitHub-hosted session working on one branch and up to one pull request per assigned task | Local repository, dedicated worktree, or GitHub-hosted cloud sandbox (Public Preview) | Configured Azure resources, telemetry, connectors, and operational controls | IDE extensions for developer execution; Modernize CLI for portfolio assessment, planning, and orchestration |
| Planning and Autonomy | Can research a repository, create a plan, change code, and optionally open a pull request | Interactive, Plan, and Autopilot modes; model and reasoning-effort selection per session | Investigation and action scope follow configured permissions and approvals | Assessment and modernization plans guide transformations; developers choose and review the work to execute |
| Parallelism and Isolation | Separate cloud-agent sessions; each task remains branch-scoped | Multiple isolated sessions in parallel, each with its own branch/worktree; cloud sandbox isolation is Public Preview | Multiple investigations can run within configured service and quota boundaries | Modernize CLI can coordinate assessment and planning across multiple applications; IDE work remains project/repository scoped |
| GitHub Workflow | Tasks can begin from issues or Copilot prompts; work is visible through commits, logs, and pull requests | Browse issues/PRs, start sessions with issue context, review diffs and CI, comment or request fixes, and use Agent Merge when GitHub rules allow | Use governed issues or incident records as the handoff to coding workflows; no direct SRE-to-coding-agent protocol is assumed | Produces reviewed code and deployment changes that should flow through normal branches, builds, tests, scans, and pull requests |
| Steering and Review | Developers manage sessions and decide whether to create or merge the pull request | Steer and archive sessions, switch modes/models, review diffs/comments, inspect CI, and use canvases where the workflow benefits from a shared custom surface | Human approval gates remain required for consequential actions | Human review remains mandatory; successful transformation does not guarantee application correctness, security, or production readiness |
| Modernization Scope | General-purpose repository work; no modernization-specific guarantee | General-purpose session and PR management | Operations, not application framework modernization | Java, .NET, and C++ language/framework/tool upgrades; supported .NET/Java Azure migration; dependency/runtime upgrades, OpenRewrite transformations, build/test validation, CVE remediation, containerization, IaC, and deployment assets |
| Feature-Level Status | Available for paid Copilot plans; related capabilities carry their own availability labels | App available for all Copilot plans; cloud sandboxes Public Preview | GA | IDE language/framework/tool upgrades GA for .NET, Java, and C++; IDE Azure migration GA for .NET and Java; Modernization agent CLI Public Preview |
Sources: GitHub Copilot cloud agent; GitHub Copilot app; agent sessions; issues and pull requests; Azure SRE Agent; GitHub Copilot Modernization (updated June 12, 2026).
Workflow Orchestration Platform Comparison
Agentic workflows automate business processes and orchestrate AI agents through deterministic sequences with AI integration. Microsoft provides three primary technologies for different development approaches and team skills.
Use this matrix to choose between Agent Framework Workflows, Logic Apps AI Agent Workflows, and Copilot Studio agent flows based on skills, hosting, connectors, and checkpointing needs.
Technology Options
| Feature | Agent Framework Workflows | Logic Apps AI Agent Workflows | Copilot Studio Agent Flows |
|---|---|---|---|
| Approach | Pro-code (C#, Python) | Visual designer + code | Low-code (natural language/visual) |
| Architecture | Graph-based (Executors + Edges) | Workflow designer + connectors | Workflow designer + connectors |
| Checkpointing | ✅ Yes (built-in) | ✅ Yes (state management) | ❌ No |
| Type Safety | ✅ Yes (compile-time) | ⚠️ Partial (schema validation) | ⚠️ Partial (schema validation) |
| AI Integration | Native (\IChatClient) | Azure OpenAI, Microsoft Foundry (Azure), Microsoft Foundry Agent Service | Azure OpenAI, Microsoft Foundry (Azure), Microsoft Foundry Agent Service |
| Connectors | Custom code (any API) | 1,400+ connectors + custom code | 1,000+ (Power Platform) |
| Hosting | Self-managed (Azure) | Azure Portal | Microsoft SaaS (managed) |
| Licensing | Open-source (free SDK) | Azure consumption/Standard | Copilot Studio Credits |
| DevOps | ✅ Yes (code-based CI/CD) | ✅ Yes (GitHub Actions, Azure DevOps) | ✅ Yes (Pipelines, GitHub Actions, Azure DevOps) |
| MCP Server | ⚠️ Custom | ✅ Yes (Preview) | ✅ Yes (Preview) |
| Status | GA | Preview (Consumption agent workflows), GA (core) | GA |
| Best For | Multi-agent orchestration + checkpointing | Enterprise integration + AI agents | Fast automation within Studio agents |
Sources:
- Agent Framework Workflows
- Logic Apps Agent Workflows
- Copilot Studio Agent Flows
- Agent Flows vs Cloud Flows FAQ
Last Updated: 2026-03-25 Confidence Level: High (official Microsoft documentation for all three technologies)
When to Use Each Workflow Technology
Agent Framework Workflows
Use When:
- Type-safe workflow orchestration with compile-time validation is required
- Checkpointing needed for long-running or human-in-the-loop processes
- Multi-agent patterns required (Sequential, Concurrent, Handoff, Magentic)
- Pro-code development team (C#, Python) available
- Integrating with M365 Agents SDK
Don’t Use When:
- Low-code approach preferred
- Need 1,000+ pre-built connectors
- Fast time-to-market is critical priority
- No development team available
Sources:
- Agent Framework Workflows Overview (Updated: 2026-02-20)
Logic Apps AI Agent Workflows
Use When:
- Enterprise integration with 1,400+ connectors (cloud + on-premises) required
- Visual designer + code hybrid approach fits team skills
- Azure-native hosting with DevOps deployments needed
- MCP server capabilities required (expose workflows as AI tools)
- Autonomous or conversational agent workflows with Azure OpenAI
Don’t Use When:
- Purely low-code approach needed (use Copilot Studio instead)
- Budget constraints prohibit Azure consumption costs
- M365-only governance required
- Compile-time type safety is critical (use Agent Framework instead)
Sources:
- Logic Apps Agent Workflows Concepts (Updated: 2026-02-19)
Copilot Studio Agent Flows
Use When:
- Native automation within Copilot Studio agents needed (no separate license)
- Fastest time-to-value with low-code/natural language approach required
- Managed SaaS preferred (no infrastructure management)
- Business process automation for Studio agents
- Makers without coding skills available
Don’t Use When:
- Checkpointing or state persistence required for long workflows
- Complex multi-agent orchestration needed
- Full DevOps control and Azure-native hosting required
- MCP server capabilities needed
Sources:
- Copilot Studio Flows Overview (Updated: 2026-03-20)
Agent Governance Layering Guide
Agent governance requires several complementary layers: runtime policy enforcement, fleet and lifecycle administration, platform guardrails, and behavioral evaluation. The guide below maps each layer to its responsibility and product anchors. It is a framework-authored composition model; Microsoft does not publish it as an official product comparison.
| Layer | Role | Product Anchor | Status / Scope | Configuration Caveat |
|---|---|---|---|---|
| Runtime checkpoint | Intercept agent actions and enforce application policy | Agent Governance Toolkit (Microsoft-origin, MIT OSS) | Public Preview v4.1.0; zero GA features; ACS policy layer beta | Self-hosted middleware; pin versions; APIs and schemas can change; no managed-service SLA or Microsoft product support contract |
| Fleet and lifecycle | Inventory, identity, access, and administrative posture | Agent 365, Agent Registry, or Foundry control surfaces | Status varies by exact surface | Registration does not automatically intercept every in-process tool call |
| Platform guardrails | Apply identity, network, data, and service policies | Hosting and cloud platform controls | Status varies by control | Coverage follows the configured service boundary, not arbitrary application code |
| Behavioral evaluation | Test whether an agent follows policy and fails safely | ASSERT and platform evaluation tools | Project/service status varies | Tests provide evidence; they do not enforce runtime decisions |
Authorship and method: This framework authored the layering model from official product scopes and the AGT repository/release record. It compares architectural responsibility, enforcement point, operational owner, and lifecycle—not feature counts. See the Microsoft announcement, AGT repository, and v4.1.0 changelog.
IQ Layer Comparison
Four intelligence layers – collectively branded Microsoft IQ – give agents access to different slices of your world. Think of them as The Four Libraries: one holds your enterprise documents, one remembers every meeting and email, one crunches your business metrics, and one reads every public source on the internet. No single librarian covers all four, but together they give an agent near-total recall.
| Feature | Foundry IQ | Work IQ | Web IQ | Fabric IQ |
|---|---|---|---|---|
| Domain | Enterprise knowledge (files, blobs, indexes, web) | M365 organizational context + extensible business systems (the intelligence layer that powers Copilot) | Public web and fresh external signals | Business analytics (semantic models, ontologies, OneLake) |
| Status | Partial GA (core knowledge source types + 2026-04-01 REST API GA; portal, answer synthesis, multi-turn remain Preview) | Preview (MCP tools and API as of June 2026) | Limited Access (enrollment required) | Preview |
| Data Boundary | Azure | M365 tenant | Limited Access web grounding service | Fabric capacity |
| Access Method | MCP endpoint (agentic retrieval via knowledge bases) | Built-in MCP servers (Copilot, Calendar, Mail, SharePoint, OneDrive, Teams, User, Word, Dataverse/Dynamics 365) + custom MCP servers via MCP Management Server (1,500+ connectors, Graph APIs, REST) | Agent-native web grounding APIs | Fabric Data Agents, SQL endpoint |
| Permission Model | ACL sync + Purview labels + Entra identity at query time | M365 Copilot license + Entra identity + admin allow/block | Limited Access terms and web-source policies | Fabric RBAC + OneLake ACLs |
| Knowledge Sources | Azure Blob, SharePoint (indexed + remote), OneLake, existing indexes, web (Bing) | M365 Graph signals (files, emails, meetings, chats, calendar, people) | Web pages, news, public information sources | Lakehouse, warehouse, KQL, semantic models |
| Consumers | Foundry Agent Service, custom agents via MCP | Copilot Studio, Microsoft Foundry, VS Code, CLI | Limited Access agent scenarios | Copilot Studio, M365 Copilot, Power BI |
| Best For | Custom agents with governed enterprise data + citation-backed responses | Agents that need organizational context: not just file contents, but who decided what, when, and what happened after | Agents needing current public information (news, research, regulations) | Chat-with-data and analytics scenarios |
| Licensing | Azure consumption | M365 Copilot license required | Limited Access enrollment | Fabric capacity (F2+) |
Sources:
- Foundry IQ overview
- Work IQ MCP overview
- Fabric IQ overview
- Microsoft IQ overview
- Web IQ product page (Limited Access; no dedicated Learn page as of June 2026)
Lifecycle & Operational Readiness Comparison
Before selecting a platform, check the expiration date. This matrix flags active migration deadlines and governance readiness.
| Dimension | M365 Copilot | Copilot Studio | Microsoft Foundry | Agent Framework | GitHub Copilot |
|---|---|---|---|---|---|
| Platform Status | GA | GA | GA | GA (.NET, Python) | GA (various features Preview) |
| Active Deprecations | None | None | Foundry Workflows: Retiring from Preview without a GA path on December 1, 2026; Assistants API (Aug 26, 2026); azure-ai-agents deprecated | Breaking changes tracked in upgrade guides | Model LTS policies active |
| Publish to M365 Copilot | Native | Native | Early Access Preview; validate each tool and identity path | Via M365 Agents SDK | N/A |
| Multi-Agent Orchestration | Via Researcher agent (GA) + connected agents | Connected agents (Preview) | Incoming A2A endpoint (Preview); use Agent Framework for code-first orchestration | Custom workflow graphs; built-in Sequential, Concurrent, Handoff, Group Chat, and Magentic-One patterns | N/A |
| MCP Compatibility | N/A | Consumer (GA) | Consumer + Producer (MCP tool GA) | Consumer (GA) | Consumer (MCP servers in VS Code) |
| Permission-Aware Grounding | Graph security trimming (automatic) | Via connectors + Work IQ MCP | Foundry IQ ACL/label sync; SharePoint ACL flow-through in AI Search | Inherits from host runtime | Repository-scoped |
| Built-in Evaluation | N/A | Test sets + analytics | Foundry portal evals + continuous evaluation | OpenTelemetry hooks | N/A |
| Governance Export | Agent Registry (export inventory) | Power Platform admin + Purview audit | Foundry Control Plane + Azure Policy + Defender | Application-level | Org admin telemetry |
| Observability | M365 admin center | Purview audit + Application Insights | Application Insights agent details view | OpenTelemetry (ENABLE_OTEL) | Usage metrics + plan mode |
| Identity Model | User-scoped (always) | Configurable (delegated or service account) | Managed identity + Entra Agent ID | Application-defined | GitHub identity |
Data Grounding Technology Comparison
Side-by-side grounding options (Copilot connectors, AI Search, Fabric, Cosmos, PostgreSQL, SQL Database Engine family) by search mode, boundary, and best-fit workloads.
| Feature | Copilot connectors | Azure AI Search | Microsoft Fabric | Cosmos DB | PostgreSQL | SQL Database Engine (Azure SQL DB, SQL MI, SQL Server 2025, SQL database in Fabric) |
|---|---|---|---|---|---|---|
| Vector Search | No (semantic index only) | ✅ Yes (IVF, HNSW) | ✅ Yes (Lakehouse via external tools) | ✅ Yes (IVF, HNSW, DiskANN) | ✅ Yes (pgvector, IVF) | ✅ Yes (DiskANN, ANN Index Preview) |
| Hybrid Search | No | ✅ Yes (vector + keyword) | ✅ Yes (SQL endpoint + external) | ✅ Yes | ✅ Yes | ✅ Yes |
| Data Boundary | M365 tenant | Azure | Azure (OneLake) | Azure | Azure | Azure SQL DB & SQL MI: Azure. SQL Server 2025: On-prem or Azure VM. SQL database in Fabric: Fabric. |
| Index Target | Microsoft Graph (connector ingestion) | Azure AI Search index / knowledge bases (agentic retrieval, preview) | Lakehouse Delta tables, Warehouse tables | Cosmos DB collection | PostgreSQL table | SQL table (VECTOR column) |
| Access Method | Graph API (Copilot connector APIs) | REST API, SDKs (2025-11-01-preview for knowledge bases) | ADLS Gen2 APIs, SQL endpoint, Fabric Data Agents | SDKs, REST API | SQL, pgvector | T-SQL, VECTOR type, VECTOR_DISTANCE, VECTOR_SEARCH |
| Best For | M365-centric knowledge (Copilot, Search, Context IQ) | Azure-native RAG with ACL/label enforcement and agentic retrieval | Analytics data + unified data platform | Transactional + vector data | PostgreSQL workloads with AI | SQL workloads with AI. Choose deployment by existing estate (cloud PaaS, lift-and-shift, on-prem, Fabric) |
| Licensing | Included with M365 | Azure consumption: dedicated tiers, plus a serverless consumption tier (Public Preview, pay-per-use, scale-to-zero); agentic retrieval preview on free tier quotas | Fabric capacity (F2+) | Azure consumption | Azure consumption | Azure SQL DB/MI: Azure consumption. SQL Server 2025: SQL license. SQL in Fabric: Fabric capacity. |
| Status | GA | GA (dedicated tiers); Serverless tier Preview; agentic retrieval preview | GA (Platform), Preview (Data Agents) | GA | GA | VECTOR type GA. ANN index (DiskANN) Preview. |
Fabric app-backend note: Rayfin is not a grounding store. It is a Preview SDK and CLI for defining and deploying managed application backends on Microsoft Fabric, so app data lands in OneLake and inherits Fabric governance. Use it when the question is “Where should this AI-enabled app backend live?” rather than “Which retrieval engine should ground this answer?”
Sources:
- Microsoft 365 Copilot connectors overview (Updated: 2026-02-25)
- Azure AI Search what’s new (Updated: 2026-03-13)
- Optimize costs for Azure AI Search Serverless (Public Preview) (Updated: 2026-06)
- Microsoft Fabric Platform (Updated: 2026-03-18)
- Rayfin product page (Retrieved: 2026-06-08)
- Microsoft Foundry (Azure) FAQ (Updated: 2026-01-23)
- Cosmos DB Vector Search (Updated: 2025-12-19)
- Azure Database for PostgreSQL AI (Updated: 2026-01-20)
- SQL Database Engine Vectors, applies to SQL Server 2025, Azure SQL Database, Azure SQL MI, SQL database in Fabric (Updated: 2026-03-18)
- Azure SQL Database AI (Updated: 2026-02-06)
- SQL database in Fabric (Updated: 2026-02-06)
Confidence Level: High for GA technologies (VECTOR type GA across SQL engine family), Medium for ANN Index (DiskANN) Preview
Azure AI Search Agentic & Hybrid Preview Highlights
| Feature | Status | Notes | API Version | Documentation |
|---|---|---|---|---|
| Knowledge bases (renamed from knowledge agents) | Preview | Routes /knowledgebases/*, outputMode replaces outputConfiguration, retrievalReasoningEffort (minimal/low/medium) replaces fast path; partial responses supported. | 2025-11-01-preview | Agentic retrieval migration |
| Knowledge sources | Preview | Indexed/remote SharePoint, indexed OneLake, web/Bing, search index, Azure Blob; ingestionParameters wrap embeddings/chat models, schedules, contentExtractionMode (Content Understanding). Portal uses 2025-08-01-preview objects. Migrate to 2025-11-01-preview. | 2025-11-01-preview | Knowledge sources |
| Content Understanding skill | Preview | Rich Markdown/table extraction and chunking without Text Split; billed to Foundry resource. | 2025-05-01-preview | Content Understanding skill |
| Semantic ranker on free tier | GA | Semantic reranking available on free tier with volume limits. | n/a | Semantic ranker |
| Hybrid/vector debug & control | Preview | truncationDimension for MRL embeddings, filterOverride for vector-only filters in hybrid queries, debug subscores for RRF, token-based Text Split parameters. | 2024-09-01-preview | Hybrid search preview |
Last Updated: January 28, 2026
Local and Edge AI Comparison
Not every model belongs in the cloud. This compact matrix helps you decide where inference should run – the datacenter, the device, or somewhere in between.
| Feature | Foundry Local | Windows AI APIs | Windows ML (ONNX) |
|---|---|---|---|
| What It Does | Runs Foundry-catalog SLMs locally with cloud-parity API | Pre-built on-device intelligence (text, image, speech) | Custom ONNX model execution on DirectML hardware |
| Status | Verify current scope on Learn | GA (Windows 11) | GA |
| Platforms | Windows 11, macOS, Azure Local | Windows 11 (Copilot+ PCs preferred) | Windows 10/11 |
| Hardware | CPU/GPU (device-class) | NPU, GPU, CPU (auto-selected) | DirectML (GPU/NPU) |
| Model Source | Foundry catalog subset (Phi, Mistral, others) | System-provided models | Bring your own ONNX model |
| Offline Capable | Yes (after model download) | Yes (built-in) | Yes |
| Best For | Hybrid cloud/edge agents, air-gapped environments, dev inner loop | Client-side features (captions, OCR, summarization) in Windows apps | Specialized vision, NLP, or sensor models in Windows apps |
| Cloud Fallback | Yes (same API targets Foundry cloud) | No (device-only) | No (device-only) |
The trade-off: Foundry Local gives you cloud portability at the cost of a narrower model catalog. Windows AI APIs give you zero-setup device intelligence but no custom model support. Windows ML gives you full model control but requires ONNX conversion and DirectML expertise. Choose based on whether you need portability, simplicity, or specialization.
Agent Development Approach Comparison
Declarative vs custom engine: when to stay with managed orchestration versus bringing your own orchestrator and hosting.
| Approach | Declarative Agents | Custom Engine Agents |
|---|---|---|
| Definition | Pre-built orchestration; configure instructions, knowledge, actions | Bring your own orchestrator (Agent Framework recommended, LangChain (Third-party)) |
| Best For | Simple → Moderate complexity; fast time-to-market | Complex workflows; multi-agent systems; custom reasoning |
| Development Model | Low-code (Copilot Studio) or Pro-code (M365 Agents Toolkit) | Pro-code only; full control over logic |
| Orchestration | Microsoft-managed orchestration (GPT-based) | You control orchestration framework and model selection |
| M365 Integration | Native M365 Copilot integration | Can integrate via M365 Agents SDK |
| Typical Timeline | Days to weeks | Weeks to months |
| Skill Level | Makers (low-code) or Developers (pro-code) | Developers required |
Sources:
- Declarative Agents for Microsoft 365 Copilot (Updated: 2025-12-01)
- Custom Engine Agents Overview (Updated: 2026-07-02)
Confidence Level: High (official Microsoft guidance)
Custom Engine Agent Tool Comparison
Tooling snapshot for custom engine agents across Copilot Studio, Teams SDK, and M365 Agents SDK (channels, orchestration fit, developer experience). Microsoft guidance positions Teams SDK and M365 Agents SDK as coexisting options with different optimization targets, not a strict replacement path.
| Tool | Copilot Studio | Teams SDK | M365 Agents SDK |
|---|---|---|---|
| Primary Use Case | Low-code custom engine agents | Teams-first collaborative agents (channels/meetings), including Bot Framework migration | Multi-channel pro-code agents with bring-your-own orchestration |
| Orchestration Options | Copilot Studio orchestration (classic topics or modern orchestrator), plus Agent Flows and connected agents for cross-agent patterns | Teams SDK (Action Planner) | Bring your own orchestration (Agent Framework recommended, LangChain (Third-party), custom planners) |
| Deployment Channels | Microsoft 365 Copilot, Teams, SharePoint, custom websites, and mobile endpoints | Microsoft 365 Copilot + Teams, optimized for collaborative Teams surfaces | 10+ channels (Teams, Slack, web chat, etc.) |
| Developer Experience | Visual designer + code | Code-first | Code-first with Toolkit in VS Code |
| Target Audience | Makers and developers | Teams app/agent developers | Professional developers and ISVs |
| Licensing Model | Per-user subscription or consumption (prepaid/Pay-as-you-go) | Included with M365 | Included with Azure Bot Service |
| Status | GA | GA | GA |
Publishing scope: Only agents built with the Teams SDK, M365 Agents SDK, or Foundry can be published to the Microsoft Commercial Store via Agents Toolkit; Copilot Studio agents are org-only by default. Source: Custom engine agents for Microsoft 365 overview.
Coexistence guidance:
- Use Teams SDK when the primary experience is collaborative and Teams-centric (channels, meetings, team conversations).
- Use M365 Agents SDK when you need broader cross-channel distribution and full control over orchestration/model stack.
Capability maturity in Copilot Studio and adjacent SDKs can shift month-to-month. Validate GA/Preview status per feature in the Copilot Studio “What’s new” feed before production commitments.
Sources:
- Custom Engine Agents for Microsoft 365 (Updated: 2026-07-02)
- Teams SDK welcome (Updated: 2026-07-07)
- Create and deploy with Microsoft 365 Agents SDK (Updated: 2025-12-02)
Confidence Level: High (all GA, official Microsoft documentation)
Related Pages
- For technology selection guidance, see Decision Framework
- For fast lookup tables, see Quick Reference
- For architecture patterns, see Implementation Patterns
Next: Quick Reference - Fast lookup tables to accompany the matrices
-
Agent Framework guidance on when not to use an agent (prefer deterministic functions/workflows). Source: Microsoft Agent Framework overview. ↩
-
Frontier Word/Excel/PowerPoint creation agents (Preview) require admin Frontier opt-in (Frontier is the early access program for experimental/preview Copilot features), Anthropic provider connection, and acceptance of Anthropic commercial terms; data is processed outside Microsoft-managed environments. Sources: Word/Excel/PowerPoint Agents, Microsoft 365 Copilot admin user access. ↩
-
Microsoft 365 Copilot release notes (Nov 25, 2025) – mobile support for custom engine agents and message-extension agents on iOS/Android. Source: Microsoft 365 Copilot release notes - Nov 25, 2025 (Retrieved: 2025-01-27). ↩
-
Agent Registry lifecycle actions (publish, activate, deploy, pin, block, remove, delete, reassign owner, export inventory). Source: Agent Registry in the Microsoft 365 admin center (Retrieved: 2026-01-23). ↩
-
The Agent Pre-Purchase Plan (P3) lets organizations buy Agent Commit Units (ACUs) in a single pool covering both Copilot Studio (metered messages) and Microsoft Foundry (Azure consumption) usage. This does not change per-platform licensing models but provides a unified pre-purchase option across platforms. Source: Agent Pre-Purchase Plan (Updated: 2026-01-15). ↩
-
Overview of the Microsoft 365 Copilot Search API (Preview) for hybrid semantic + lexical search across OneDrive via Graph
/beta. Source: Microsoft 365 Copilot Search API overview (Retrieved: 2025-10-20). ↩ -
What is Microsoft Foundry Agent Service?, Microsoft Learn. GA update includes Azure Logic Apps triggers for agents. Source: Foundry Agent Service overview (Retrieved: 2026-03-25). ↩
-
Workflows with AI agents and models in Azure Logic Apps (Preview), Microsoft Learn. Source: Logic Apps agent workflows concepts (Retrieved: 2026-02-19). ↩
-
What is Microsoft Foundry Agent Service?, Microsoft Learn. MCP tool available as built-in agent tool. Source: Foundry Agent Service overview (Retrieved: 2026-03-25). ↩
-
Set up Standard logic apps as remote MCP servers (Preview), Microsoft Learn. Source: Set up Standard logic apps as remote MCP servers (Retrieved: 2025-11-18). ↩