You have been building skills one model at a time. Now you learn to use them in concert. The sophistication of the architecture should follow the complexity of the problem, not the other way around.
What did your verification workflow reveal? When you rebuilt a high-stakes task with explicit checkpoints and tried a step in a different model, what did the process catch that you would have missed?
The goal is not whether the second model agreed with the first. The goal is whether the comparison changed your confidence in the result, up or down.
Listen for stories where the verification step caught something meaningful. Those stories become the room's evidence that multi-model thinking is worth the extra effort. If nobody found anything, that is worth discussing too: was the original output actually solid, or did the verification step not go deep enough?
A quick orientation to the five multi-agent patterns, not a walkthrough. The full reference is on the Multi-Agent Patterns page, which was assigned as a pre-read.
The question to put to the room: which of these patterns have you already been doing without naming them? Anyone who has copied output from one model and pasted it into another has been running a relay. Anyone who has asked Claude to argue against its own recommendation has been running a debate loop. The patterns are vocabulary for behavior you have already started.
Most valuable work today lives in Patterns 1 through 3. Start with a single agent. Add a relay when you feel the limits of one context window. Add a debate loop when you need adversarial pressure. The sophistication of the architecture should follow the complexity of the problem.
Walk through the email processing system as a design case study. This is not about email. It is about how to think through an AI system where the value comes from organizing your attention rather than automating your judgment.
Focus on three design decisions that teach principles applicable to any system you might build:
Trust calibration. The config file assigns draft confidence per contact. Strategic relationships stay at "low" because the cost of a wrong tone exceeds the time saved by a draft. That is the same judgment call you make every time you decide how much to trust AI output.
Intentional friction. The system drafts but never sends. Removing the human from the loop is not the goal. Removing the noise between the human and the decision is the goal.
Boundary rules. The system refuses to draft for unknown contacts. It does not guess at intent for first-contact emails. Notice what the system does not do. That is a design decision, not a limitation.
Click through the simulation and let the room react to each email. The Mon Mothma placeholder draft versus the C-3PO auto-response is the contrast that makes the trust calibration principle concrete. Ask: where in your own work would you set the threshold at "low" versus "high"?
Each participant picks a real challenge from their work and designs a multi-model workflow on paper. Not in Claude. On paper or out loud. The thinking is the exercise, not the implementation.
Choose a pattern from the five. Assign the roles: which model does what and why. Define the handoffs: what moves from one step to the next, and what gets checked at each transition. Identify where the human stays in the loop and where the system can move on its own.
Share your design with the room. The group pressure-tests: is the pattern too simple for the problem? Too complex? Where would it break first?
This exercise is the bridge to the group project. The workflow you design here might become the candidate challenge you bring to Module 7. Think about problems that are worth solving together, not just individually.
The final three modules of the program are a collaborative build. Your cohort will select a shared challenge, design the solution using the patterns and principles from the first six modules, and present the result to all three cohorts at the capstone session.
Module 7 is definition, design, and delegation. The cohort selects one challenge from the candidates everyone brings, scopes it, assigns roles, and produces a project brief.
Module 8 is build and review. Working session to execute, troubleshoot, and prepare the presentation.
Module 9 is the capstone. All three cohorts present their projects to each other in a combined 90-minute session. This is the cross-cohort moment that has been building since the beginning.
Use this time to surface candidate problems from the room. What is a problem in your work or your organization that would benefit from the skills you have built in this program? Something concrete enough to build a solution for in two sessions. Something worth showing to 15 other executives. Get the ideas on the table and start evaluating them together.
The best project candidates are real problems, not hypothetical ones. They are scoped tightly enough to build in two sessions but meaningful enough that the result is worth presenting. They apply at least one of the patterns from this module. And they produce something tangible: a working prototype, a designed system, a decision framework that others could adopt.
Your homework is to take the ideas generated in this discussion and prepare to execute in Module 7. That means sharpening your problem statement, sketching the system design, and thinking about how the work divides across the group. Module 7 is a working session, not a brainstorming session. Arrive ready to produce a project brief.
You can identify which multi-agent pattern fits a given problem and explain why you would choose it over the alternatives. You can design a multi-model workflow with explicit handoffs, trust calibration, and human checkpoints. You understand that the value of orchestration is not in the automation but in how it organizes your attention and judgment.
Most importantly, you have candidate project ideas on the table. Your homework is to prepare for Module 7 by sharpening your problem statement, sketching the design, and thinking about how to divide the work. Module 7 is where the brief gets written.