CAF phases meet the agent lifecycle
A unified model aligns the Cloud Adoption Framework (CAF) phases with the agent lifecycle to reduce risk, prevent “agent sprawl,” and accelerate value — carrying a workload systematically from proof of concept to production:
Strategy → Plan agents
Document business outcomes, rank use cases by impact/feasibility, decide if agents fit; produce an AI strategy brief and technology plan.
Plan → PoC
Convert strategy into an adoption plan; run a proof of concept and define agent-readiness criteria (data, governance, identity, connectors).
Ready → Govern & secure (foundation)
Build Azure landing zones, an agent-governance charter, and a data-access model on authoritative sources.
Govern + Secure AI → enforcement
Apply policy-based governance (Azure Policy), agent behavior/data controls, risk registers, and audit trails.
Build agents → execution
Standardize development on Copilot Studio/Foundry with templates, evaluation gates, and CI/CD guardrails.
Manage AI → operate agents
Run the fleet with monitoring, SLOs, cost controls, and an Agent Ops playbook (retraining, deprecation).
Choose the shape, then the orchestration
Default to a single agent for bounded domains; go multi-agent only when you must cross security/compliance boundaries, multiple teams own distinct knowledge, the roadmap spans 3–5+ functions, or work has many cross-workstream dependencies. Assign the right platform to each role:
| Platform | Role | Best for |
|---|---|---|
| Microsoft 365 Copilot | Domain agents embedded in M365 (SaaS) | Immediate value, limited customization |
| Copilot Studio | Task & retrieval agents (low-code) | Business-led processes, moderate customization |
| Microsoft Foundry | Connected agents & orchestration (pro-code) | Strategic, complex workflows, custom tools |
Key concept · Five orchestration patterns
- Sequential — deterministic pipeline (Plan → Enrich → Verify → Act)
- Concurrent — parallel agents on independent subtasks, then reconcile
- Group chat — agents propose, a moderator arbitrates (consensus)
- Handoff — transfer context/control to a specialist (or human) on escalation
- Magentic — dynamic runtime selection of the right expert agents
Connected agents take clear roles — Planner, Researcher, Reviewer, Actuator — and orchestration is treated as a workflow with state, branching, and error handling, not a brittle prompt-to-prompt chain.
The core design decision
Extend Microsoft 365 Copilot when the scenario is productivity-shaped, the data lives in M365, custom logic is simple, and Microsoft's built-in Responsible AI guardrails suffice. Build a custom agent when you need specialized multi-step workflows, external system APIs, multi-agent collaboration, or execution outside M365.
| Attribute | Copilot extension | Custom agent |
|---|---|---|
| Autonomy | Low | High |
| Custom logic | Limited | Extensive |
| Data variety | Microsoft 365 | Any enterprise data |
| Actions | Simple | Complex, multi-step |
| Governance | Built-in | Custom required |
The module also covers the surrounding design choices: when to create custom AI models or use small language models (SLMs), choosing generative-AI knowledge sources for Copilot Studio agents, building a prompt library with prompt-engineering guidelines, defining solution rules and constraints, identifying key business user roles, evaluating regional data-regulation compliance, composing multi-app Dynamics 365 solutions, and designing user prompt training for adoption.
The AI Center of Excellence
An AI Center of Excellence (CoE) is a strategic function that accelerates responsible AI adoption, standardizes practice, and prevents ungoverned “shadow AI.” Its essential elements: executive sponsorship, a single accountable CoE leader, and a multidisciplinary team (AI strategy, data, ML, governance, security, MLOps, and business SMEs). It owns strategy, skills enablement, governance/standards, project intake & prioritization, delivery support, and operations.
Key concept · CoE maturity — and its anti-patterns
A CoE evolves through three stages: Centralized (full control, approves projects) → Hybrid (CoE keeps standards, product teams take ownership) → Advisory (CoE consults; governance embedded in platforms). Watch the anti-patterns: overcentralization → bottlenecks; undergovernance → shadow AI; no executive sponsor → stalled adoption; the CoE acting as a gatekeeper instead of an enabler.
Module 3 quiz
Fifteen questions on CAF adoption, multi-agent design, the build-vs-extend decision, and the CoE. Answer first, then expand Show answer.
Part A · Multiple choice
What risk does aligning CAF phases with the agent lifecycle specifically help prevent?
- A Model overfitting
- B “Agent sprawl” and architecture drift
- C Token overuse
- D Data duplication
Show answer
B. The unified model reduces risk and prevents agent sprawl, moving workloads systematically from PoC to production.
In the CAF-to-lifecycle mapping, running a proof of concept and defining readiness criteria happens in which phase?
- A Strategy
- B Plan
- C Manage
- D Build
Show answer
B — Plan. Strategy documents outcomes and ranks use cases; Plan converts that into an adoption plan and validates with a PoC.
Which orchestration pattern has agents propose ideas that a moderator agent arbitrates?
- A Sequential
- B Concurrent
- C Group chat
- D Magentic
Show answer
C — Group chat. It's for consensus-building. Magentic dynamically pulls in the right expert agents at runtime.
Which platform is the pro-code choice for connected agents and complex orchestration?
- A Microsoft 365 Copilot
- B Copilot Studio
- C Microsoft Foundry
- D Power Automate
Show answer
C — Microsoft Foundry. M365 Copilot is SaaS domain agents; Copilot Studio is low-code task/retrieval agents.
Which is a signal to build a custom agent rather than extend Copilot?
- A Data lives entirely in Microsoft 365
- B Simple summarization of documents
- C Multi-step orchestration calling external system APIs outside M365
- D You want Microsoft's built-in Responsible AI guardrails to do the work
Show answer
C. Custom agents suit specialized multi-step workflows, external integrations, multi-agent collaboration, and execution outside M365.
Which are the four roles a set of connected agents typically take?
- A Planner, Researcher, Reviewer, Actuator
- B Vision, Speech, Language, Decision
- C Strategy, Plan, Ready, Manage
- D Build, Buy, Extend, Retire
Show answer
A — Planner, Researcher, Reviewer, Actuator. Each connected agent gets a single clear responsibility and a minimum toolset.
Part B · True or false
The module recommends starting with multi-agent designs and only simplifying to a single agent later.
Show answer
False. Start simple with a single agent; scale to multi-agent only when specific evidence (security boundaries, multiple owners, broad roadmap) requires it.
Orchestration should be treated as a workflow with state, branching, and error handling.
Show answer
True. That's what makes multi-agent systems reliable, versus brittle prompt-to-prompt daisy chains.
Extending Microsoft 365 Copilot inherits Microsoft's built-in Responsible AI guardrails and compliance.
Show answer
True. Extension provides enterprise-grade safety and built-in compliance; custom agents require you to build your own governance.
A mature AI Center of Excellence acts primarily as a gatekeeper approving every project.
Show answer
False. Maturity moves Centralized → Hybrid → Advisory, where the CoE consults and governance is embedded in platforms; acting as a gatekeeper is an anti-pattern.
Part C · Short answer
Name the six CAF phases in the unified adoption model.
Show answer
Strategy, Plan, Ready, Govern (+ Secure) AI, Build agents, and Manage AI — mapped onto plan → govern & secure → build → operate.
List the five orchestration patterns.
Show answer
Sequential, Concurrent, Group chat, Handoff, and Magentic.
Give three conditions that justify a multi-agent (rather than single-agent) design.
Show answer
Any three of: crossing security/compliance boundaries; multiple teams owning distinct knowledge or release cycles; a roadmap spanning 3–5+ functions; work with many cross-workstream dependencies.
What is an AI Center of Excellence, and name three of its essential elements.
Show answer
A strategic function that accelerates responsible AI adoption and standardizes practice. Three elements (of many): executive sponsorship, a dedicated CoE leader, and a multidisciplinary expert team (strategy, data, ML, governance, security, ops, business SMEs).
Name two CoE anti-patterns to avoid.
Show answer
Any two of: overcentralization (bottlenecks), undergovernance (shadow AI), no executive sponsor (stalled adoption), or the CoE acting as a gatekeeper rather than an enabler.