AB-620 cheat sheet
One page for the last review before the exam. The study guide explains the reasoning, and the open-book strategy covers what to look up during the exam.
Pass = 700 · INTEGRATE 40–45% · PLAN 30–35% · MANAGE 20–25% · Default = generative orchestration
1. Exam technique
| Habit | Fix |
|---|---|
| Rushing | Take ~45 s. Strike unlimited / never / always / only / nothing-optional first |
| Ordering items | Prerequisite first. For each step ask: what must already exist? |
| Multi-select second pick | Test it against every stem constraint: no API → no connector; retrieval ≠ model choice; no "convert" step |
2. Sequences
| Setup | Order |
|---|---|
| App Insights | resource → connection string (Settings › Advanced) → enable logging → publish |
| AI Search | add knowledge → Azure AI Search → connection → auth + endpoint → index → add |
| REST wizard | add tool → spec + solution → auth → operations → connection → test |
| Agents SDK | publish → enable channel → get config → implement SDK → validate → prod |
| Responsible AI | DLP → content moderation → prompt modifications |
3. Easy to get wrong (know instantly)
| Cue | Answer |
|---|---|
| Connected agent actions | publish → callable · Add → collaboration · Configure connections → invocation. An agent is never a knowledge source |
| No API, must automate a UI | Enable Computer use; watch the test pane's live remote session (5 credits/step, 15 premium) |
| Governed model for a prompt | Foundry model catalog setting in the prompt, never an instruction line |
4. Choose the integration
| Need | Choose |
|---|---|
| Indexed third-party content | Copilot connector |
| Live LOB data, no index | Connector as real-time knowledge |
| Write back to a system | Connector tool |
| Tool server that changes | MCP: tools + resources, generative only |
| OpenAPI spec | REST tool: blank solution = auto-created; fill every parameter description; per-user → OAuth 2.0 |
| No API at all | Computer use: UI-only, autonomous |
| Agent on another platform | A2A: sends full chat history; works in a VNet |
| Agents SDK agent | Activity Protocol |
| Foundry agent | Project endpoint + Agent ID |
| Fabric data agent | No redirect · no M365 Copilot · no evaluation |
| Same team builds it | Child agent |
| Other team, own ALM or model, reuse | Connected agent: can't nest; inherits general knowledge OFF; citations might not survive; ~30–40 choices before splitting |
5. Knowledge
| Fact | Value |
|---|---|
| Websites, generative / classic | 25 / 4 (SharePoint 25 generative) |
| Dataverse knowledge | 2 sources × 15 tables |
| More than 25 sources | Filtered by description; uploaded files excluded |
| Azure AI Search | 1 index per connection · URL field for citations · private endpoint → Power Platform VNet |
| SharePoint file size | 200 MB (512 MB for PDF/PPTX/DOCX) |
| Ungrounded answers OFF | Uncited answers are withheld → add citation instructions |
| Classic-only | "Official source" · Bing Custom Search → Classic data node |
6. Orchestration and identity
- Descriptions beat trigger phrases. Agent decides → generative orchestration; fixed step → tool node
- Ignored under generative: Conversational boosting, Multiple Topics Matched · custom entities can't be inputs
- End all topics = cancel the plan · Reset Conversation ≠ clear history
- Teams/M365 → Authenticate with Microsoft · need
User.AccessToken→ manual auth · no auth = anyone with the link - Auth changes need a publish · D365 Customer Service handoff ✗ Authenticate with Microsoft
7. Flows and topics
- Trigger When an agent calls the flow + Respond to the agent · failing flow → check run history first
- Cloud flow → agent flow is one-way · multistage approval = AI + manual + conditions, no ALM
- Capacity: topic = classic answer · orchestrator = autonomous action · test pane = free
- HTTP node: headers for auth, 30 s timeout, Continue on error
- Generative answers: disclaimers go in the prompt modification · clear Send a message for cards · node moderation wins
- Regex case-insensitive =
(?i)· Secrets → env var → Key Vault
8. Monitor and evaluate
- App Insights:
customEvents·designMode == "False"· anonymous users = random ID per conversation · block via data policy - Graders: General quality (no expected answer) · Compare meaning (50) · Exact · Similarity · Tool use · Keyword · Safety (0 = strict) · Custom for business rules · no expected answer → Invalid
9. ALM
- Custom solution → add agent → export (managed is an export option; "convert first" is fake)
- Env vars → pipelines → targets must be Managed Environments
- Delegated deployments = approval + service principal · no cross-tenant · no stage skip · Upgrade import
10. Harness (new)
Standard = rule-based (this exam) · GitHub Copilot = goal/files/skills/memory · Copilot chat = extends M365 Copilot Chat