Module 3 · Lesson 6 of 8
Venue research and catering research involve different questions. One needs seating, location, and accessibility information. The other needs menus, dietary requirements, and per-person prices.
We could give one agent both jobs. We could also separate them into two agents, each with focused instructions and suitable tools.
Separate responsibilities can help
A multi-agent system coordinates more than one agent. The agents do not have to use different underlying models. Two instances of the same model can have different tasks, tool access, and context.
In our example, the venue agent investigates spaces while the catering agent investigates food. A coordinator brings their findings together and prepares a proposal for the organizer.
Give a task, not just a role name
“You are the venue expert” leaves several questions unanswered. A useful assignment says: find up to three available venues for 40 guests on the requested date, report accessibility details, include price sources, and do not make reservations.
The catering agent needs the same date and guest count. It also needs to know whether a venue might require an in-house caterer. Two separate searches can produce individually attractive options that cannot be used together.
Useful agent boundaries therefore follow the dependencies in the work. If the venue determines every catering choice, it may be better to select a venue first.
Delegation and handoff are different arrangements
With delegation, the coordinator asks a specialist to do a subtask and receives its result. The coordinator still owns the overall response. With a handoff, another agent takes over the active responsibility, such as continuing the conversation about a reservation.
These two arrangements appear as agents-as-tools and handoffs in the OpenAI Agents SDK documentation on Agent orchestration.
Share evidence with the conclusion
“Riverside is best” gives the coordinator little to inspect. A better result includes the venue, quoted price, source, date checked, requirements met, and unresolved questions.
Sharing every message from every agent is not always helpful either. Each worker needs the relevant task information. The final decision needs the evidence behind the recommendations.
More agents introduce more coordination
Two agents can repeat the same search, interpret a requirement differently, or both assume the other checked an extra fee. The coordinator needs a way to recognize incomplete and conflicting results.
Parallel work can reduce elapsed time when tasks are independent, but it also uses resources across several workers. Shared model behavior and shared sources can produce shared mistakes; agreement is not independent verification.
For an account of real delegation, parallel research, and coordination tradeoffs, see How we built our multi-agent research system.
Use multiple agents when separating the work has a clear benefit. Role names alone do not make the result better. Measure whether the team improves the task compared with a simpler system.