Copyright (c) 2026 MindMesh Academy. All rights reserved. This content is proprietary and may not be reproduced or distributed without permission.

5.2. Quick Reference

🎯 Decision: Which knowledge source?

The scenario saysAnswer
"External content searchable like SharePoint / alongside M365"Copilot connectors (ingestion into semantic index)
"Answers must reflect live records / current state"Power Platform connector as knowledge
"We already have an index" / "full retrieval control" / "VNet-isolated corpus"Azure AI Search (vector index, integrated vectorization)
"Answers must use our Azure OpenAI deployment on our index"Azure OpenAI connection on the generative answers node

🎯 Decision: Which tool type?

The scenario saysAnswer
"A governed/custom connector already exists"Reuse the connector
"API spec, only this agent needs it"Add REST API directly
"New operations must appear without republishing" / "shared, evolving toolset"MCP tools
"No API exists — UI only"Computer use (generative orchestration + dedicated Windows machine)

🎯 Decision: Which multi-agent mechanism?

The scenario saysAnswer
"One team, grouping tools/intents, routing degraded (30–40+ tools)"Child agent
"Another team owns it / reused by several parents / separate ALM"Connected Copilot Studio agent
"Built by developers in Foundry / fine-tuned model"Connected Foundry agent
"Natural-language analytics over lakehouse/warehouse/KQL"Fabric data agent (no topic redirect; not M365 Copilot)
"Different vendor/platform/organization"A2A protocol (agent card discovery)

🎯 Decision: Which evaluation method?

Correct answer is…Method
One exact string (code, number)Exact Match
Must contain specific termsKeyword Match
Same wording/structure expectedText Similarity
Same meaning, any phrasingCompare Meaning
Open-ended, no expected answerGeneral Quality
"Did it use the right tool/topic?"Tool Use
Your own rubric/labelsCustom

🎯 Decision: Identity and channels

AudienceAuthNatural channels
EmployeesMicrosoft Entra ID (user context; SSO)Teams, Microsoft 365 Copilot
Known external usersOAuth provider (manual config)Custom website, custom app
Anonymous publicNo auth — public content onlyCustom website, messaging
No human (autonomous)Service/maker connections + supervisionEvent/schedule triggers

🎯 Canonical rules that decide questions

  • Node-level data sources override agent-level knowledge (fallback only).
  • Variables are topic-scoped by default; global by explicit promotion; system variables are platform-set.
  • Flow invoked by an agent needs the "When an agent calls the flow" trigger; error handling is authored via run-after conditions.
  • Default solution can't ship managed — build in a custom solution; per-environment values live in environment variables (text / data source / secret); credentials rebind via connection references.
  • Approval before prod = pipeline extension; unattended deployment = service principal.
  • App Insights needs its connection string configured — nothing flows by default.
  • Azure AI Search knowledge: vector index with integrated vectorization, one index per connection, URL/metadata_storage_path field for citations.
Alvin Varughese
Written byAlvin Varughese
Founder18 professional certifications