Design a multi-agent workflow around a shared task.

Define the shared outcome

Describe the final artifact and the constraints it must meet. A marketing brief might need an audience hypothesis, approved facts and three creative directions. Give every participant the same source of truth so effort does not split into incompatible answers.

Separate independent work

Research, drafting and quality review can be useful roles when their boundaries are clear. Avoid assigning several agents the same vague task and hoping their outputs agree. Give each role a deliverable that another person can inspect.

Make handoffs explicit

Specify the input format, decision owner and condition for moving forward. A reviewer should know which claims need sources and which choices are open. If a prerequisite changes, update the shared brief rather than passing corrections informally through several layers.

Limit authority

Treat external pages and messages as untrusted inputs. Give tools only the permissions needed for their assigned step. Consequential actions should remain governed by the human owner’s authorization and a record of what was approved.

Review the combined result

Check contradictions, missing sections and repeated claims after merging the work. Measure the total time including coordination and repair. A single focused assistant can be the better option for a small task with tightly coupled decisions.

Split the work where responsibilities actually differ

Identify parts of the task with different evidence or quality criteria. Research, drafting and checking may need different instructions, but three copies of the same broad request can produce duplicated work rather than useful specialization. Give each role a narrow responsibility and an explicit completion condition.

Start with a single-task baseline so you can compare the added coordination cost. If the task is small or tightly coupled, separate roles may introduce more confusion than they remove. The design should explain why the boundary exists and which dependency crosses it.

Create a shared brief and a handoff contract

Record the objective, approved facts, audience, exclusions and output format in one shared reference. A role should receive the information it needs without silently rewriting the goal. At handoff, include the deliverable, source notes, unresolved questions and the checks already performed.

Keep ownership clear when two outputs conflict. The workflow needs a person or designated review step that decides what enters the final result. Agreement between several generated drafts is not independent verification of a factual claim. Check important claims against their actual evidence.

Constrain tools and information by role

A drafting role may only need approved notes, while a publishing role would involve a different level of authority. Do not grant every role all available access simply to make the setup symmetrical. Define which actions remain drafts and which require a human decision.

Make stop conditions visible: missing evidence, contradictory requirements, unexpected data or unavailable dependencies. A role should return an explicit issue rather than inventing a completed step. This helps the coordinator tell the difference between useful progress and an unsupported assertion of success.

Evaluate the combined output and the coordination burden

Inspect the complete result for inconsistent terms, repeated sections and gaps between handoffs. Track the review effort needed to assemble it. A workflow can produce individually polished pieces while still failing as a coherent deliverable.

For the academy exercise, compare the multi-role version with the baseline using the same brief and review criteria. Record where specialization helped and where it created rework. Scale the arrangement only if the improvement is visible. More agents are a design choice, not a quality metric or a substitute for accountability.