Use this cheat sheet for Microsoft Applied Skills: Create an AI agent as a lab-readiness checklist. The assessment is not a broad certification exam. It evaluates whether you can create and configure agents in Microsoft Foundry, add knowledge sources and tools, and orchestrate multiple agents.
Agent build sequence
flowchart TD
S["Scenario goal"] --> P["Create Foundry project and agent"]
P --> I["Write instructions and behavior constraints"]
I --> K["Add knowledge source"]
K --> T["Add tools or actions"]
T --> O["Orchestrate multiple agents if required"]
O --> V["Test expected, edge, and failure cases"]
If the lab task asks for an agent, do not start by adding every possible tool. Start with the agent purpose, instructions, and evidence that it behaves correctly.
What to know cold
| Lab area |
Decision rule |
Reject when |
| Agent purpose |
Define the job, audience, limits, and expected output. |
The agent has vague instructions and cannot be tested. |
| Model/deployment |
Select a model/deployment that fits the lab task and environment. |
You troubleshoot instructions while the model/deployment is missing. |
| Knowledge |
Add the correct source and test whether the agent uses it. |
Uploaded material is assumed useful without grounded tests. |
| Tools |
Add only tools needed by the task and validate inputs. |
The agent can call broad tools with no constraints. |
| Orchestration |
Split work across agents only when roles or steps justify it. |
Every task is forced into one giant prompt. |
| Validation |
Test normal, missing-data, and unsafe-input paths. |
The agent passes one happy path and is treated as complete. |
Lab-task eliminations
| Stem clue |
Eliminate first |
Keep in play |
| “create and configure” |
only open the portal |
project, model/deployment, agent, instructions, test |
| “add knowledge” |
paste all facts into prompt |
attach source, index/connect it, and test grounded answers |
| “add tools” |
broad unvalidated action |
scoped tool, clear schema, input validation, test result |
| “orchestrate multiple agents” |
one agent does all work |
role-specific agents with clear handoff |
| “answer from source” |
model’s general knowledge only |
knowledge source and citation/grounding behavior where available |
| “wrong or unsafe output” |
change random wording |
tighten instructions, source scope, tool constraints, and validation |
Foundry agent checklist
| Step |
Evidence that it is done |
| Agent exists |
Name, description, instructions, and model/deployment are configured. |
| Instructions are usable |
The agent has task, boundaries, style, and refusal/clarification behavior. |
| Knowledge is connected |
The agent can answer from the intended source and fails gracefully when source evidence is missing. |
| Tools are configured |
Tool schema, permissions, and inputs match the task. |
| Multi-agent path works |
Each agent role is clear and the handoff produces a useful result. |
| Tests pass |
Expected, edge, and failure cases are checked before finishing. |
Final 15-minute review
| If the lab asks for… |
Start with |
| “configure an agent” |
purpose, model/deployment, instructions, and behavior constraints |
| “knowledge source” |
source connection, retrieval test, and grounded answer |
| “tool” |
exact action, schema, permission, validation, and result check |
| “multiple agents” |
role separation, routing, handoff, and end-to-end test |
| “troubleshoot output” |
instructions, knowledge retrieval, tool result, and test evidence |