Quick takeaways
- Teach agents through a real workflow, not abstract definitions.
- Every workshop should end with a concrete artifact.
- Scope and review gates are part of the lesson.
What are AI agents?
Microsoft Developer explains what AI agents are, when to use them, and how to start building one safely.
Workshop agenda
This agenda fits a 90-minute session. Shorten exercises for a 60-minute version.
- Introduction (10 min): Define agents, prompts, and automations. Clarify that agents are not magic.
- Pick a workflow (15 min): Each group selects one repeated task with clear inputs and outputs.
- Scope the task (15 min): Write the job, constraints, success criteria, and failure mode.
- Map tools and review gates (20 min): List tools the agent can use and where humans must approve.
- Draft instructions (20 min): Write the system prompt and step-by-step plan.
- Test and debrief (10 min): Walk through one imaginary run and identify risks.
Exercise set
Scope the task
Write the job, constraints, and expected output. Be specific enough that another team could reproduce it.
Map the tools
List the tools the agent can inspect or use. Start read-only.
Define review
Set the human approval point before any risky action.
Draft the prompt
Write the system instructions, expected steps, and output format.
Materials
- Printed worksheet with sections for workflow, tools, review gates, and prompt.
- Example agent designs from support triage, research, and code review.
- A list of common failure modes to review before testing.
- Access to the AI agents explained guide for reference.
Artifacts to leave with
- One scoped agent idea.
- One permission and review model.
- One prompt or instruction set.
- One test or evaluation checklist.
- One follow-up plan with owner and deadline.
Follow-up actions
The workshop is only useful if the ideas turn into action. Within one week, the team should:
- Assign an owner to the top agent idea.
- Build a minimal prototype in a safe environment.
- Run three to five test cases and document failures.
- Schedule a review to decide whether to expand or scope down.
Without follow-up, workshop artifacts become shelfware. Make the first prototype small enough to finish quickly and concrete enough to evaluate.
Without AI vs. with AI
| Task | Without AI | With AI |
|---|---|---|
| Scoping | Workshop groups start with a vague idea of what the agent should do. | Teams define a bounded goal, inputs, success criteria, and stopping conditions. |
| Safety | Discussions ignore permissions, approvals, and failure modes. | Teams list allowed tools, required review gates, and escalation paths. |
| Demo | The workshop produces a one-off prototype that no one can reproduce. | The output is a reusable agent spec with prompts, tools, and evaluation. |
| Handoff | Ideas stay on sticky notes and whiteboards. | Teams leave with a documented runbook, prompts, and a 30-day test plan. |
| Adoption | Participants leave inspired but unsure what to build first. | Each team picks one concrete, low-risk internal workflow to pilot. |
FAQ
How long should the workshop be?
60 to 90 minutes is enough if you stay focused on one workflow.
Who should attend?
Founders, operators, developers, or marketers who own the workflow you want to improve.
Do participants need coding skills?
No. The workshop focuses on design, scope, and safety. Technical implementation comes later.
How many people per group?
Three to five people works well. Smaller groups force participation; larger groups slow down.
What is the most important outcome?
A concrete, scoped agent design with clear review gates. Avoid leaving with vague "AI ideas."
Can this workshop be run remotely?
Yes. Use a shared document or whiteboard tool so all participants can contribute in real time.
How long should an AI agents workshop be?
Sixty to ninety minutes is enough to scope one safe agent and leave with a concrete plan.
What should we build during the workshop?
Pick an internal workflow with clear inputs and outputs, not a customer-facing or production-changing one.
How do we keep the workshop from becoming theoretical?
End with a written spec, assigned owners, and a two-week test window with success metrics.