FIELD BRIEFING // TARGET: EXAM AB-620 // EYES ONLY: LEVI
Recon complete /// objective assessed

Operation AB-620

Microsoft's AI Agent Builder Associate cert. The official courseware is a swamp. This is the map through it — built on the fact that you already own a third of the ground.

Real title
Build integrated agents in Copilot Studio
Level
Intermediate
Time
120 min
Pass mark
700 / 1000
Questions
~40–60
Fee
USD 165
Proctor
Pearson VUE
Perk
Split-screen MS Learn

Territory already controlled

0% of exam weight

Tick items as you lock them in. The green ones are pre-checked because you do them for a living — that's your head start.

Phase 00 // salvage

Stuff you already know, wearing a Microsoft nametag

Half the misery of this exam is Microsoft re-explaining agents to someone who ships multi-agent systems on opencode and runs MCP servers by hand. Don't relearn these. Just learn the branded noun and move on.

OWNED

MCP tools — it's literally the Model Context Protocol. You run MCP servers. Copilot Studio just gives you a wizard to point an agent at one. Learn the 3-click flow, not the protocol.

OWNED

Custom connectors / Send HTTP request / REST API tools — REST + OAuth 2.0 you already do in bug-bounty and tooling. This is the same, in a form designer.

OWNED

Grounding / knowledge sources — it's RAG. Retrieval you already understand. The only new part is which Microsoft product does the retrieving (Azure AI Search).

OWNED

Multi-agent orchestration + A2A protocol — you build agent hand-off already. A2A is just the open standard for it. Connected/child agents = your sub-agents with a GUI.

OWNED

Custom prompts / instructions — prompt engineering. Nothing new here except the box you type it in.

OWNED

Application Insights monitoring — telemetry / observability. Familiar territory, Azure branding.

Phase 01 // the actual target

Three domains. Where your study hours actually go.

Each domain below maps Microsoft's termwhat you already know, then flags the genuinely-new bits and the traps they'll set in the questions. Tap a domain to open it.

Plan the solution

  • Identity strategy who authenticates: end-user vs agent auth, SSO, which connector auth model. // scenario picking
  • Channels & deployment Teams, a website, M365 Copilot, external. Internal vs external audience changes your design.
  • Responsible AI strategy content moderation, groundedness, guardrails. They want the Microsoft-named feature, not your opinion.
  • Security & governance DLP policies, environment strategy. // your AD/Entra brain helps here
  • Reusable components over rebuilding each time.

Agent flows NEW

  • Agent flows Power Automate–style automations sitting behind the chat agent.
  • Human-in-the-loop a step where a person approves before the flow continues.
  • Configure actions & connectors, add input/output parameters, implement error handling.

Configure topics NEW UI

  • Trigger phrases & the authoring canvas — how a topic fires.
  • Generative answers node let the model answer from knowledge vs an authored topic path.
  • Adaptive cards JSON-defined rich UI cards for responses. New — learn the shape.
  • Power Fx variables Excel-formula-flavored expression language. Global vs topic scope. Memorize the syntax feel.
  • Custom knowledge sources, API & Send HTTP request actions.
⚠ Trap radar

Generative answers node vs authored topic — they hand you a scenario and make you pick. Rule of thumb: unpredictable / open-ended → generative; deterministic / compliance-critical flow → authored topic.

Variable scope — global (whole agent) vs topic-local. Easy points if you know it, easy loss if you don't.

Enterprise knowledge sources

  • Copilot connectors (Graph connectors), Power Platform connectors — pull in org data.
  • Azure AI Search the vector/RAG index service. // RAG = owned, product = new

Tools & orchestration

  • MCP tools you run these already
  • REST API tools / custom connectors REST+OAuth, owned
  • Computer use agent takes UI actions on the user's behalf. Newish — know when it's the right tool.

Multi-agent collaboration

  • Connected / child agents agents delegating inside Copilot Studio. // your sub-agents
  • A2A protocol agent-to-agent hand-off across platforms/vendors.
  • Foundry agent integration & Fabric data agent (querying enterprise semantic models). New names, learn what each is for.

Azure integration

  • Generative answers via Azure AI Search + Foundry.
  • Foundry model catalogue for custom prompts.
  • Monitor with Application Insights. // observability, owned
⚠ Trap radar — this is where points bleed

MCP tool vs custom connector vs REST API tool — they give a scenario, you pick the mechanism. Speaks MCP → MCP tool. Existing REST endpoint you define → custom connector / HTTP action. Don't overthink it, but know the tell in each question.

A2A vs connected agents — same platform = connected/child agent. Different vendor/platform = A2A. This distinction shows up constantly.

Foundry agent vs Copilot Studio agent — know why you'd reach into Foundry (model catalogue, pro-code models) vs staying native.

Evaluate performance NEW-ish

  • Create a test set, choose an evaluation method, review results.
  • It's QA for agents — prove behavior before it ships. Microsoft treats it as first-class.

ALM — application lifecycle management NEW — DON'T SKIP

  • Solutions the container you package an agent in (managed vs unmanaged).
  • Environment variables so nothing's hardcoded across dev/test/prod.
  • Power Platform Pipelines promote a solution dev → test → prod without manual rebuild.
⚠ Trap radar

This whole domain is the sleeper. Coming from pure red-team/dev work, ALM (solutions, env variables, Pipelines) is the least-familiar and the cheapest points to grab. Do one hands-on lab moving a solution between environments and this domain becomes free marks.

Phase 02 // pressure test

Recon drills

The exam is mostly "given this scenario, pick the mechanism." These are the recurring decisions. Answer in your head, then reveal.

ScenarioYour agent must call a third-party tool that already speaks MCP. What do you add?

An MCP tool — point the agent at the MCP server. Not a custom connector, not a raw REST action. If it speaks MCP, use the MCP path.

ScenarioYour agent must hand a task to an agent running on a different vendor's platform. Which protocol?

A2A (Agent2Agent). Cross-platform hand-off = A2A. If both agents lived in Copilot Studio, you'd use connected/child agents instead.

ScenarioYou need the same agent to move from dev to prod with no hardcoded URLs or keys. What's the play?

Solution + environment variables + Power Platform Pipelines. Package in a solution, parameterize with environment variables, promote via a Pipeline. This is Domain 3's core.

ScenarioGround the agent in a large enterprise document corpus with semantic/vector search. Which service?

Azure AI Search, surfaced as generative answers (often via Foundry). It's RAG — you know the concept; here they want the Azure product name.

ScenarioA flow must pause for a manager to approve before it continues. What are you configuring?

A human-in-the-loop agent flow with an approval step. Domain 1, agent flows strand.

ScenarioWhere do you write formula logic to set or transform a variable's value?

Power Fx. Excel-formula-flavored. Get comfortable with the syntax feel — it shows up across topics and flows.
Phase 03 // execution

The compressed study plan

The exam rewards hands-on and punishes crammers — it's built around scenarios you can't fake. So the plan is mostly labs, weighted toward Domain 2.

Get a live environment

Spin up a Power Platform trial, open Copilot Studio, build and publish one working agent end-to-end. Non-negotiable — everything else builds on this.

day 0

Live in Domain 2 (the 42%)

Lab, don't read: connect Azure AI Search grounding, add a REST API tool, wire an MCP tool, set up connected agents + an A2A hand-off, integrate a Foundry agent. This is most of your hours.

bulk of your time

Clear Domain 1's new bits

Build an agent flow with a human approval + error handling. Author a topic using the generative answers node and an adaptive card. Touch Power Fx variables until the syntax stops feeling foreign.

the branded UI

Grab the ALM free marks

Create a solution, add your agent, use an environment variable, push it through a Power Platform Pipeline dev→test. One lab converts the sleeper domain into easy points.

don't skip

Drill responsible-AI & naming

Memorize the Microsoft-named features for moderation, groundedness, governance. Pure-dev candidates lose cheap points by describing the concept instead of naming the feature.

cheap points

Exam-day mechanics

You get split-screen Microsoft Learn — pre-bookmark the ALM, Foundry, and connector reference pages for quick lookups. It's a safety net for one detail, not a substitute for the labs. Clock keeps running.

go time
Phase 04 // supply drop

The three official learning paths (free)

These are the real spine of the exam and they're free. Use them as the outline; do the labs above for the actual retention. Skim the parts you already own.