Turning Three Conflicting Agent Proposals Into One Decision
I give talks and workshops on this material for retail teams and conferences. See past talks or invite me to speak.
Series: Foundations of Agentic AI for Retail, Chapter 8 of 14. Updated for the Third Edition (October 2026).
From the book: Foundations of Agentic AI for Retail, Third Edition
Thursday morning, midway through a Monday-to-Sunday promotion. In the chapter's running case, one store has 25 units of a promoted product, SKU-123, left against 40 forecast for the rest of the week. Three agents read the same numbers.
The pricing agent proposes ending the discount on Friday. The inventory agent proposes a 15-unit transfer from a nearby store, arriving Friday, so the advertised price can hold. The customer service agent wants to offer a substitute if the shelf empties.
Each recommendation makes sense on its own, and the first two contradict each other. Ending the discount breaks an advertised price, and the transfer only helps if the units arrive by Friday.
The job of a team of agents is to turn those proposals into one decision record that its owner can approve.
Why strong agents can still fail as a team
The proposals can be compared only if the agents share definitions. Chapter 8 starts with four words teams use interchangeably, which mean different things in production. Here is what each one means on SKU-123's Thursday.
| Word | What it means for SKU-123 |
|---|---|
| Coordination | The price and the transfer must not undo each other |
| Orchestration | A supervisory agent routes the price proposal for approval |
| Collaboration | All three agents read the same stock evidence |
| Delegation | The transfer is handed off with a clear test: arrive by Friday |
Agents need shared interfaces too, so the chapter builds the team on explicit contracts instead of ad hoc messaging.
Without them, one agent adds a required field to its message, another still expects the old version, and the handoff fails quietly. In the section "Contract-First Multi-Agent Systems," the verdict is blunt: "The system appears healthy but decisions are missing."
What should reach the category manager's screen
Here the category manager owns the decision, and any change to an advertised price inside the promotion window needs that manager's approval.
The chapter opens a decision record for SKU-123 on its first page and carries it to the last. It names the owner and the rule, weighs the transfer against holding the price without one, and keeps a list of open evidence. One open item is whether all 25 units are really on the shelf, since Chapter 7's moved product shows how a count can mislead.
The book's capstone fills that record with an orchestrator-worker-audit loop, shaped like this.
Because the price proposal changes an advertised price, the orchestrator sends it to the category manager. That last step is where many systems come apart.
Multi-agent systems often break at this final mile, where an agent must present a decision to a human.
— Foundations of Agentic AI for Retail, Chapter 8
By the last page, the record holds an owner's decision, its evidence, and a fallback in case the transfer misses Friday.
When a team beats one agent, and when it does not
Of all the projects in my courses at McGill, a multi-agent system is the one my students found most interesting and learned the most from.
My own team at OODARIS AI supplies the counterweight. We have built several agents, with orchestration and handoffs, where one agent with the right tools would have done. We have also had an agent grow so large that we had to divide it into agents and subagents.
The chapter's coordination mechanisms carry the same judgment. If one donor store can confirm the units and the Friday arrival, the chapter delegates the transfer to it directly. Contract Net, where agents bid for a task, is for when several donors compete.
What the chapter gives you
- The orchestrator-worker-audit loop, the capstone's way of bringing several agents' recommendations to one owner
- SKU-123's decision record, followed from Thursday's three proposals to the category manager's decision
- Four collaboration patterns, each with a job and a failure to test on SKU-123, plus coordination mechanisms from Contract Net to auctions
- Runnable companion demos, from agent messaging to a supplier auction, and exercises that build up to designing SKU-123's agents
Go deeper
Chapter 8 of the Third Edition, out since October 2026, has the full method, the SKU-123 case, and the companion code. It also covers what this post only hinted at: "Capstone Integration: Minimal Orchestrator-Worker-Audit Loop" and "Coordination Mechanisms," which weighs Contract Net, markets, and consensus. New in the Third Edition: Part III follows one incident that several agents act on, and asks who owns it.