RESEARCH REPORT PART IV APR 2026 ARUP DIGITAL ADVISORY TECHNICAL

Building custom
agents, seriously.

A technical reference on the Microsoft agent-development stack — Azure Foundry, Agent Framework, Copilot Studio, M365 Agents Toolkit and SDK. For team members ready to move beyond Agent Builder and ship production-grade custom agents. Prerequisites up front.

Who this is for
You're already comfortable in Python or .NET. You've built something with an API before. You know what git clone means. You've outgrown Agent Builder and want to know what's next. This report is the map.
Foundry
Model catalogue
spans OpenAI, Anthropic, Meta, Mistral, more
100+
Prebuilt Copilot connectors
(SAP, Salesforce, ServiceNow…)
v1.0
Microsoft Agent Framework
production release
5 patterns
Multi-agent orchestration
shipped with MAF 1.0
01 / PREREQUISITES

Before you start — what you need.

The tools in this report are powerful because they're real developer tools. That also means they assume a working knowledge of cloud, code, and command line. Read this first. If more than two items in the Skills column feel unfamiliar, spend a week on the Learning Path in Section 9 before going further.

The advanced-tools prerequisite checklist.
Set up · Skills · Certifications · Access

1 · Environment setup

  • Active Azure subscription Personal or Arup-provisioned. A free Azure account includes a trial subscription with $200 credit for experimenting.
  • Visual Studio Code The default IDE for Python, JS/TS, and declarative agent work. Install the Microsoft Foundry extension and the Python extension.
  • Visual Studio 2022+ (if .NET) Required for Microsoft 365 Agents Toolkit for .NET and any C# agent work.
  • Python 3.10+ Minimum 3.9 supported, 3.10+ recommended. Always use a virtual environment — never install SDK packages into global Python.
  • Node.js 18+ Required for the M365 Agents Toolkit MCP Server and for JS/TS agent development.
  • .NET 8 SDK Required only if you're building in C# — also applies to some Microsoft 365 Agents SDK templates.
  • Azure CLI + Azure Developer CLI az login and azd — both used heavily across all four tools in this report.
  • Git & GitHub account Every sample ships from GitHub. If you don't use Git today, make this the first skill you pick up.

2 · Skills you need

  • Python or C#/.NET You don't need both. Pick one. Python is more common in the AI ecosystem; .NET is common across Arup's existing stack. Both are fully supported.
  • REST APIs & SDK usage Calling an HTTP API, reading a JSON response, authenticating with a token. This is the day-to-day reality of agent work.
  • Async programming basics Agents are async by nature. async / await, streaming responses, event-driven flows. Intimidating at first, fundamental by week two.
  • JSON and YAML Declarative agents and YAML-based agent definitions are everywhere. No code required to read them, but you'll live in them daily.
  • Basic cloud concepts What a subscription, resource group, and role assignment are. What managed identity means. What RBAC does. Not deep Azure — just literate.
  • Microsoft Entra ID awareness Every Microsoft agent tool uses Entra for auth. Know what an app registration is. Know what DefaultAzureCredential means in practice.

3 · Access & permissions

  • Azure RBAC — development Azure AI User (least privilege) on the Foundry projects you use, or Azure AI Project Manager to manage projects.
  • Azure RBAC — setup Owner on the subscription or resource group is needed to create Foundry resources for the first time.
  • Microsoft 365 Copilot licence Required to publish declarative agents to M365 Copilot surfaces (Word, PowerPoint, Chat).
  • Teams Developer Portal access For building and sideloading Teams-surface agents. Your admin may need to enable developer mode.
  • Entra app registration rights The Agents Toolkit will register Entra apps for you during provisioning. Admin consent may be required for Graph permissions.

4 · Certifications path

  • AI-900 / AI-901 (Azure AI Fundamentals) Foundational. AI-901 is the updated Azure AI Fundamentals exam with stronger Foundry + generative AI coverage; AI-900 remains available alongside it. Check Microsoft Learn for current retirement dates.
  • AI-102 (Develop AI solutions in Azure) The next rung. Assumes C# or Python + REST API experience. Covers generative AI, computer vision, NLP, and information extraction on Azure.
  • Microsoft Learn free courses The "AI Agents for Beginners" course (12 lessons) covers Agent Framework, AutoGen, Semantic Kernel, and Foundry Agent Service end-to-end.
  • Sample repos first, then certifications Do two end-to-end samples before sitting any exam. Certifications confirm what you already know; they don't teach it.
Plan data residency early
Foundry hosted agents are available across a growing list of Azure regions — East US 2, North Central US, Sweden Central, UK-adjacent regions via virtual networks, Australia East, and more. Foundry projects span an even wider set of regions including UK South. Always check the current hosted-agent region list and your subscription quota before deploying for client work, and plan data-residency implications into any project that uses Foundry Agent Service.
02 / LANDSCAPE

Five tools, one map.

Microsoft's advanced agent stack looks fragmented until you understand where each tool sits. Here they are in the order of increasing technical depth — from low-code Copilot Studio through open-source Agent Framework to full production on Foundry.

You've met Agent Builder in the prior report — the no-code interface inside the M365 Copilot app, great for small team helpers grounded on a SharePoint folder. This report is about everything past that point. It's where you go when Agent Builder's boundaries start to pinch — when you need custom actions, multi-step flows, external API calls, multi-agent orchestration, or production deployment.

The five tools below aren't competitors — they're a ladder. You'll often use two or three together on the same project. The question isn't "which one" but "which one first."

Tool Skill floor Publish where Best for
Copilot Studio (advanced) Low-code. Visual designer + YAML view. No mandatory code. Teams, Web, WhatsApp, SMS, Copilot Chat, M365 Copilot Department or external-facing agents with connectors to SAP / Salesforce / Dynamics / custom APIs.
M365 Agents Toolkit (declarative) JSON, YAML, light TypeScript. VS Code extension. M365 Copilot (Chat, Word, PowerPoint), Teams Pro-code agents for M365 Copilot. Runs on the same M365 Copilot orchestrator as Researcher and Analyst — new Copilot features tend to land in declarative agents first.
M365 Agents SDK (custom engine) Full-stack Python, JS/TS, or .NET. M365 Copilot, Teams, Web, Email, SMS, any channel Custom-engine agents. Bring your own AI / orchestration. Multi-channel.
Microsoft Agent Framework Python or .NET. Open-source SDK. Anywhere — local, Azure, AWS, GCP, or Foundry Multi-agent orchestration. Sequential, concurrent, handoff, group-chat, Magentic-One patterns.
Microsoft Foundry (Agent Service) Python, .NET, JS/TS, Java. Managed runtime. Any Foundry endpoint. Distribute to M365 Copilot, Teams, Entra Agent Registry. Production agents at enterprise scale. Extensive Foundry model catalogue, private networking, Foundry IQ (RAG), full observability.
The name changed — don't get confused
Azure AI Foundry was rebranded to Microsoft Foundry in late 2025. You'll see both names across docs, blogs and old bookmarks. They're the same product. Similarly, Teams Toolkit became Microsoft 365 Agents Toolkit. Same product, broader scope.
03 / COPILOT STUDIO

Copilot Studio — low-code, serious power.

The natural next step after Agent Builder. Still low-code enough for a non-developer, but with the connectors, flows, and governance to support department-wide and external-facing agents. This is where most technically-capable advisors will spend their time.

Low-code Power Platform 1,500+ connectors YAML view
Microsoft Copilot Studio

Copilot Studio is a standalone web portal (and the big sibling of Agent Builder). It targets IT pros, solution architects and power users — not developers by trade, but people comfortable with configuration, logic trees, and API connectors. It's included with your Microsoft 365 Copilot licence for internal agent use, with pay-as-you-go Copilot Credits for external deployment and advanced features. Every agent you build can be promoted from Agent Builder into Copilot Studio without starting over.

Core capabilities
  • Visual conversation designer with topics, entities, variables
  • 1,500+ prebuilt connectors (SAP, Salesforce, Dynamics, Jira, ServiceNow, etc.)
  • Custom connectors for any public API — OAuth + OpenAPI supported
  • Power Fx expressions for logic; Adaptive Cards for rich UI
  • Power Automate flow integration for multi-step workflows
  • Multi-agent orchestration — agents that call other agents
  • UI automation (RPA-like) for legacy systems with no API
  • YAML code view for programmatic control of the agent definition
  • Four agent types: Q&A, workflow, autonomous (event-triggered), cross-system
Governance & lifecycle
  • Application Lifecycle Management — dev / test / prod environments
  • Connector-level governance (admins block risky connectors)
  • Environment-level DLP policies, role-based access, auditing
  • Microsoft Purview integration for sensitivity labels
  • Publishing controls — admin approval for org-catalog publish
  • Detailed usage analytics with generative vs non-generative breakdown
  • Managed via Power Platform admin centre (not M365 admin centre)
  • Microsoft Agent 365 integration (GA May 2026) for enterprise agent governance
When to choose it
  • The agent will serve a whole department, not just your team
  • You need connectors to external systems (SAP, Salesforce, etc.)
  • You need conditional logic, branches, and variable handling
  • External-facing deployment (customer website, WhatsApp, SMS)
  • Enterprise governance is a hard requirement from the client
  • You want a visual designer over writing code
When it's the wrong tool
  • You need multi-agent orchestration patterns (use Agent Framework)
  • You need access to the full Foundry model catalogue
  • The agent is tightly coupled to M365 Copilot chat (use Agents Toolkit)
  • You want a deep custom engine with full programmatic control
  • Cost-sensitive at scale — Credit consumption adds up quickly
04 / M365 AGENTS TOOLKIT

Agents Toolkit & SDK — pro-code, M365-native.

The pro-code path for building agents that run inside M365 Copilot — and the closest thing Microsoft has to "real developer" tooling for Copilot surfaces. VS Code extension, declarative manifests, one-click provisioning. If you can write TypeScript or Python and know JSON, you're ready.

Pro-code VS Code extension M365 Copilot orchestrator JSON · YAML · TypeScript · Python · .NET
Microsoft 365 Agents Toolkit

Agents Toolkit is a suite of tools (VS Code extension, Visual Studio extension, CLI) for building declarative agents for M365 Copilot and custom-engine agents via the M365 Agents SDK. Templates, scaffolding, one-click provision/debug, CI/CD actions for GitHub and Azure DevOps, plus the Microsoft 365 Agents Playground for local testing without a dev tenant. Declarative agents built here run on the M365 Copilot orchestrator — the same orchestrator that powers Researcher, Analyst and other first-party Microsoft agents — which means new Copilot capabilities tend to land in declarative agents before Copilot Studio gets them.

Declarative agents (JSON + YAML)
  • Extend M365 Copilot's personality, instructions, knowledge, actions
  • Published to Word, PowerPoint, Copilot Chat, Teams
  • Ground on SharePoint, OneDrive, public websites, Graph connectors
  • Add MCP server actions via Add Action → Start with an MCP server
  • Use Adaptive Cards for rich response semantics and citations
  • Shared M365 Copilot orchestrator means new M365 features tend to land here first
Custom-engine agents (M365 Agents SDK)
  • Bring your own AI model and orchestration logic
  • aiohttp hosting (Python) or ASP.NET Core hosting (.NET)
  • AgentApplication routing, streaming, Copilot Studio client
  • Multi-channel publish: M365 Copilot, Teams, Web, Email, SMS
  • Integrates with Agent Framework for orchestration
  • Full control over the agent loop and tool invocation
# The declarative-agent workflow in a nutshell # 1. Scaffold a new agent VS Code → Microsoft 365 Agents Toolkit → Create a New Agent → Declarative Agent # 2. Add an MCP action to connect a data source Add Action → Start with an MCP server → paste endpoint + OAuth config # Toolkit auto-wires manifest.json, ai-plugin.json, declarativeAgent.json # 3. Provision + debug locally Provision in the Lifecycle pane → Start debugging → sideloads into M365 Copilot # 4. Ship Distribute via the Copilot agent store (Partner Center validation) or Deploy in-org via the Microsoft 365 admin centre
Toolchain highlights
  • M365 Agents Playground — test without a dev tenant or tunnel
  • Agents Toolkit MCP Server — bridges GitHub Copilot Agent Mode into the toolkit so you can scaffold agents conversationally
  • Environments (dev/test/prod) via .env files
  • Bicep infrastructure-as-code with smart defaults for Azure
  • One-click Provision + Start debugging for rapid iteration
  • Built-in CI/CD actions for GitHub and Azure DevOps
Prerequisites specific to this toolkit
  • VS Code (or Visual Studio for .NET)
  • Node.js 18+ for the toolkit and MCP server
  • Teams Developer Portal access for sideloading
  • Entra app registration rights (toolkit handles creation)
  • BOT_ID and SECRET_BOT_PASSWORD env vars configured
  • Partner Center account (only if publishing to the public store)
The declarative agent secret
The advantage most teams miss: declarative agents run on the same M365 Copilot orchestrator as Microsoft's first-party agents (Researcher, Analyst, Knowledge). That means you're inheriting the deep reasoning, work-data grounding, and advanced capabilities Microsoft ships for its own agents — without building any of it. Copilot Studio agents run on a different orchestrator, and typically get the same features later.
05 / AGENT FRAMEWORK

Microsoft Agent Framework — the engine.

The open-source SDK that converges AutoGen (Microsoft Research multi-agent patterns) with Semantic Kernel (enterprise foundations). Production-ready at v1.0 as of early 2026. This is where advanced teams build sophisticated multi-agent systems.

Open source v1.0 — production Python · .NET MIT licence MCP + A2A
Microsoft Agent Framework

Agent Framework is the open-source foundation for building sophisticated multi-agent systems. It hit v1.0 in early 2026 after several months in release candidate. You install it with one line in Python or .NET, and you're running agents in under 20 lines of code. But the reason it matters isn't simplicity — it's that it brings the five orchestration patterns from Microsoft Research (sequential, concurrent, handoff, group chat, Magentic-One) into a single production-ready library with streaming, checkpointing, human-in-the-loop approvals, pause/resume, OpenTelemetry, and Entra ID authentication.

# Install (Python) pip install agent-framework # A minimal agent from agent_framework import ChatAgent from azure.identity import DefaultAzureCredential agent = ChatAgent( name="BIMAssessor", instructions="Score BIM maturity against ISO 19650.", credential=DefaultAzureCredential(), ) # Add tools, memory, handoffs to another agent # Ship to Foundry Agent Service as a hosted runtime — one line
Orchestration patterns (all v1.0 stable)
  • Sequential — agents run in a fixed pipeline, output-to-input
  • Concurrent — agents run in parallel, results are merged
  • Handoff — agents pass control with context when they hit their scope limit
  • Group chat — agents converse with each other until consensus or resolution
  • Magentic-One — the generalist multi-agent pattern from Microsoft Research, with a planner coordinating specialists
  • All patterns support streaming, checkpointing, human-in-the-loop, pause/resume
What makes it enterprise-ready
  • Declarative YAML for instructions, tools, memory, topology
  • MCP (Model Context Protocol) for dynamic tool discovery
  • A2A (Agent-to-Agent) for cross-runtime agent collaboration
  • OpenTelemetry for observability; Entra ID for auth
  • Migration assistants from AutoGen and Semantic Kernel
  • GitHub Copilot SDK + Claude Code SDK harness integrations
  • Cross-cloud portability: Azure, AWS, GCP, or local
  • Production customers: KPMG Clara AI, Sitecore, Elastic
The KPMG bellwether
KPMG Clara AI — KPMG's global audit platform used across a regulated industry — is built on Microsoft Agent Framework and Foundry Agent Service. When a Big 4 firm chooses a framework for its flagship audit product, that's a signal. MAF is production-grade enough for global enterprise deployments today.
06 / FOUNDRY

Microsoft Foundry — the platform.

The production platform for enterprise-grade AI. Unified portal, managed agent runtime, extensive Foundry model catalogue, broad connector ecosystem, private networking, full observability. Where the agents built in every other tool on this list ultimately run in production.

Production platform SDKs GA March 2026 80% of Fortune 500 Python · JS/TS · Java · .NET
Microsoft Foundry Agent Service

Microsoft Foundry (renamed from Azure AI Foundry in late 2025) is a unified platform combining a portal, an agent runtime, and a model catalogue. 80,000+ enterprises use it, including 80% of the Fortune 500. Foundry Agent Service — the runtime core — went GA in May 2025, is built on OpenAI's Responses API, and is wire-compatible with OpenAI agents. Foundry SDKs (Python, JS/TS, Java, .NET) all shipped stable 2.0.0 releases in March 2026. This is where production-grade agents actually live.

Platform capabilities
  • Foundry model catalogue — Anthropic's Claude family, OpenAI (GPT-5 / GPT-5.2, Codex), Cohere, DeepSeek, xAI, Meta, Mistral and more
  • Foundry IQ (public preview) — dynamic RAG reasoning via Azure AI Search, multi-source grounding, iterative retrieval, reflection. 3B+ queries/day
  • Foundry Agent Service — managed runtime with Entra-based agent identity, RBAC, content safety, versioning, publishing
  • Multi-agent workflows — visual builder in the portal; persistent state, error recovery, context sharing
  • Agent-to-Agent (A2A) tool (preview) — call any A2A-protocol endpoint
  • Foundry MCP Server (preview) — cloud-hosted MCP at mcp.ai.azure.com with Entra auth
  • Memory in Foundry Agent Service (preview) — managed long-term memory with automatic extraction/consolidation/retrieval
  • Voice Live — managed real-time speech-to-speech runtime
  • Broad enterprise connector ecosystem via Azure Logic Apps and Copilot connectors (SAP, Salesforce, Dynamics, ServiceNow…)
Enterprise production features
  • Private networking — BYO VNet, no public egress, container/subnet injection
  • Foundry Control Plane — consolidated ARM API for governing agents, models, tools
  • Agent identity — each agent gets its own Entra identity, scoped credential-free access
  • Fine-grained RBAC — control who can create, invoke, and manage agents
  • Content safety — integrated filters for prompt injection and unsafe outputs
  • Full tracing + evaluation — eval results linked to agent traces
  • Versioning + rollback — automatic snapshots; promote to stable endpoint on publish
  • Distribute to M365 Copilot, Teams, Entra Agent Registry from one platform
  • Foundry Local — private preview on Android for edge deployment
  • GA integrations with Microsoft Defender and Purview for compliance
# The typical Foundry Agent Service development loop from azure.identity import DefaultAzureCredential from azure.ai.projects import AIProjectClient project = AIProjectClient( endpoint="https://resource.ai.azure.com/api/projects/my-project", credential=DefaultAzureCredential(), ) # 1. Build — define an agent with instructions, tools, model # 2. Test — chat in the agents playground or run locally # 3. Trace — inspect every model call, tool invocation, decision # 4. Evaluate — run evals to catch regressions between versions # 5. Publish — promote to a managed resource with a stable endpoint # 6. Monitor — track performance via the Agent Monitoring Dashboard # Added in April 2026: Task Adherence (preview), LangGraph integration, # Computer Use tool, Image Generation tool, SharePoint tool, # Microsoft Fabric data agent (preview)
When Foundry is the right answer
  • Client is deploying agents at scale — hundreds or thousands of users
  • Hard requirements for private networking, data residency, or sovereign deployment
  • Need the latest frontier models (Claude, GPT-5.2, specialist reasoning models)
  • Production SLA matters — uptime, monitoring, tracing are non-negotiable
  • Multiple specialised agents need to coordinate (use with Agent Framework)
  • Integration with enterprise systems via the broad connector catalogue (Logic Apps + Copilot connectors)
When it's overkill
  • You're still in the "does this idea work at all" prototyping phase
  • The agent only needs to serve a small Arup team internally
  • Simple Q&A over SharePoint is enough — use Agent Builder
  • The work is bounded to M365 Copilot surfaces — use Agents Toolkit
  • Budget is zero — Foundry is serious infrastructure, priced accordingly
Azure ML SDK v1 deprecation
If any existing Arup project still uses Azure Machine Learning SDK v1, migrate it now. End of support is 30 June 2026, after which workflows may face security risks and breaking changes with no Microsoft support. The AzureML CLI v1 extension already reached end-of-support on 30 September 2025. The SDK v2 migration guide is the starting point — v2 brings YAML-first job definitions and continued investment.
07 / DECISION TREE

Which tool, when?

A five-question decision tree for picking the right advanced agent tool. Walk through it top to bottom. Stop at the first question where the answer is clear.

The advanced-tool decision tree.
Five questions · walk them in order · stop at first clear answer
01
Is this a small helper for my team on an existing SharePoint folder?
Yes → Don't use anything in this report. Use Agent Builder from Part III and ship in 15 minutes.
No → continue.
02
Will the agent live primarily inside M365 Copilot (Chat, Word, PowerPoint)?
Yes → Use M365 Agents Toolkit with declarative agents. Same M365 Copilot orchestrator as Researcher and Analyst, JSON/YAML, one-click provision from VS Code. Best pro-code path for M365-native agents.
No → continue.
03
Does the agent need to integrate with systems outside Microsoft 365 (SAP, Salesforce, custom APIs, ServiceNow)?
Yes, and low-code is fine → Copilot Studio. 1,500+ connectors, visual designer, YAML view for advanced control.
Yes, and I need full programmatic control → continue to Q4.
No → revisit Q2 — you're probably still in M365.
04
Do I need multiple agents working together — sequential pipelines, handoffs, group chat, Magentic-One?
Yes → Microsoft Agent Framework. Open-source SDK, Python or .NET, all five orchestration patterns, A2A + MCP, cross-cloud portable.
No, single agent is enough → continue.
05
Am I ready to ship this to production for a real client, at scale, with SLAs?
Yes → Microsoft Foundry Agent Service. Managed runtime, broad Foundry model catalogue, Foundry IQ for RAG, private networking, full observability. Distributes to M365 Copilot / Teams / Entra Agent Registry.
No, still prototyping → build in Agent Framework locally. Migrate to Foundry when it's time to ship.
The multi-tool reality
Most real production projects use two or three of these tools together, not one in isolation. A common pattern: build multi-agent orchestration in Agent Framework, publish to M365 Copilot surfaces via Agents Toolkit, and run the whole thing on Foundry in production. Copilot Studio is the odd one out — use it alone for department-wide low-code deployments, or not at all.
08 / ARUP SCENARIOS

Where advanced tools actually unlock value.

Four realistic Arup advisory scenarios where Agent Builder simply isn't enough, and the full advanced stack pays for itself. Each scenario shows the exact stack.

Scenario 01 · Phoenix-scale deployment
An Arup-wide "Red Thread" agent for a global practice area
A senior advisor wants to build an agent that embodies the full Arup red-thread methodology for transport infrastructure advisory — consulting across 10,000+ past engagements, reading policy docs, scoring client briefs against the methodology, and being used by consultants in every region. This is Phoenix-scale: thousands of users, multi-region, SLA-grade.
Stack Foundry Agent Service Foundry IQ (RAG) Microsoft Agent Framework M365 Agents Toolkit (publish)

Flow: MAF for the multi-agent orchestration (retrieval → scoring → synthesis). Foundry IQ for grounded retrieval across the 10k past engagement corpus. Deployed on Foundry Agent Service with private networking and RBAC. Published to M365 Copilot via Agents Toolkit so consultants invoke it from Chat, Word, or Teams. Observability via Foundry tracing + evals. Governance via Agent 365 (May 2026).
Scenario 02 · SmartBid augmentation
Automated bid-response agent with deep connector integration
An advisor wants a bid-response agent that pulls structured data from SmartBid, references past winning proposals in SharePoint, checks the client in Salesforce CRM, and generates a first-draft bid section scored against Arup methodology. Needs connectors beyond M365 and departmental roll-out.
Stack Copilot Studio Custom connector (SmartBid API) Salesforce connector Power Automate flows

Flow: Copilot Studio agent with a custom connector for SmartBid (OpenAPI spec). Out-of-box Salesforce connector for client lookup. Multi-step Power Automate flow handles the sequencing. SharePoint as knowledge source for proposal precedents. Deployed to Teams for the bid team, with dev/test/prod environments and Purview sensitivity labels. No custom engine needed — this is exactly what Copilot Studio exists for.
Scenario 03 · Domain-specific ISO 19650 agent
A BIM maturity-assessment agent built as a declarative agent for M365 Copilot
A BIM specialist wants an agent that lives in the Word ribbon and Copilot Chat for advisors on BIM engagements. It knows ISO 19650 deeply, can reference the client's existing BIM execution plan, scores maturity on the spot, and drafts findings directly into the current Word document the advisor is writing. Must feel native to M365 Copilot.
Stack M365 Agents Toolkit Declarative agent (M365 Copilot orchestrator) SharePoint knowledge MCP action for BIM scoring

Flow: Declarative agent built in VS Code with the M365 Agents Toolkit. ISO 19650 PDFs + Arup BIM exec plans as SharePoint knowledge sources. A custom MCP server (hosted on Azure Functions) exposes a "score maturity" tool the agent calls. The shared M365 Copilot orchestrator gives the agent the same reasoning backbone as Researcher. Ships to Word + Copilot Chat. This is where declarative agents shine over Copilot Studio — native M365 feel plus always-latest capabilities.
Scenario 04 · Multi-agent stakeholder analysis
Orchestrated stakeholder/power-interest mapping system for a combined-authority programme
A programme advisor wants an orchestrated system for stakeholder mapping on a mass-transit programme: one agent scrapes public register and news for stakeholder background, one agent interviews the project team via Teams meetings, one agent analyses interest/power, one agent drafts the mapping report. Four specialised agents working together, handing off, with human-in-the-loop approval at each stage.
Stack Microsoft Agent Framework Magentic-One pattern Foundry Agent Service (hosting) MCP tools (news, register lookup) Teams connector

Flow: MAF with the Magentic-One pattern — a coordinator agent plans the analysis and delegates to four specialists. Pause/resume and human-in-the-loop approvals between stages. Tools exposed via MCP (one for register lookups, one for news sentiment, one for Teams transcripts). Hosted on Foundry Agent Service for production. Researcher for the final competitive-landscape section. Output lands in a Copilot Page the team edits live.
The pattern in all four scenarios
No advanced scenario uses just one tool. Copilot Studio stands alone when the requirement is department-wide external-system integration with low-code. Everything else — the Phoenix-scale, the multi-agent orchestration, the deep M365 native — involves two to four tools composed together. Advisory team members picking up these tools should plan to learn at least two of them, not one.
09 / LEARNING PATH

A six-week on-ramp.

The fastest practical route from "comfortable with M365 Copilot" to "shipping a production agent." Assumes no prior Azure or agent experience, but Python or .NET basics. Four stages.

WEEK 1
Foundation
~6 hours
Azure free account. Install VS Code + Python 3.10 + az login + azd. Walk through the "AI Agents for Beginners" course (lessons 1–4). Aim: one working local agent that calls Azure OpenAI.
WEEKS 2–3
Agent Framework
~16 hours
Install agent-framework. Build a single-agent sample with tools. Then a two-agent handoff. Then a Magentic-One multi-agent system. Clone the microsoft/spec-to-agents event-planner sample and run it locally.
WEEK 4
M365 Agents Toolkit
~10 hours
Install the VS Code extension. Scaffold a declarative agent. Add SharePoint knowledge and an MCP action. Provision, debug locally, ship to your own Copilot Chat. You now have one agent on each of the two most important surfaces.
WEEKS 5–6
Foundry + ship
~14 hours
Create a Foundry project. Deploy one of your week 2–3 agents to Foundry Agent Service. Wire up Foundry IQ for RAG on a corpus. Add tracing + one eval suite. Publish a stable endpoint. Take AI-901 (optional) to validate.

After six weeks, you'll have: a local MAF multi-agent system you wrote, a declarative agent that ships in M365 Copilot, a Foundry-hosted production agent with tracing, and enough grounding to read every Microsoft agent blog with comprehension. That's the bar for "ready to build for a client."

"Building AI agents shouldn't be rocket science. Yet many developers find themselves wrestling with complex orchestration logic, struggling to connect multiple AI models, or spending weeks building hosting infrastructure just to get a simple agent into production." — Microsoft Agent Framework launch, Oct 2025

The tools in this report exist precisely to stop that from being the bottleneck. Six weeks of deliberate practice is enough to move past Agent Builder and be genuinely useful to a client engagement that needs custom agents.

Playbook implication
The advisory team will only have a handful of members who cross into this territory — and that's fine. The playbook's job isn't to train everyone, it's to make sure those who want to can find the path, and those who don't can still brief clients intelligently on what's possible.