quiz · 4 min

Multi-Agent Systems Check

When should you reach for a multi-agent architecture instead of a single agent with tools?
Whenever parallel execution would reduce task latencyOnly when roles are genuinely distinct and cannot share a single instruction without becoming incoherentWhenever the task requires more than three sequential tool callsWhen multiple agents need to write to the same shared resource
In the supervisor/worker pattern, what is the supervisor's responsibility?
Executing the work of each subtask directly to keep latency lowSharing a single instruction set with all workers to avoid duplicationWriting all results to shared state so workers can read each other's outputRouting subtasks to workers whose roles match and assembling the results
What makes a handoff well-designed according to the lessons?
It passes the entire conversation history so the worker has full contextIt allows the worker to re-route back to the supervisor if the task is ambiguousIt includes the supervisor's full reasoning so the worker can double-check the routing decisionIt provides a short context plus a clear, narrow, self-contained instruction
How should cost budgeting work in a multi-agent system?
Each worker tracks its own step count and halts independently when its budget is exhaustedCost is the same as for a single agent because workers share the supervisor's contextBudget total steps across the whole team, not per agent, because cost compounds across agentsCost only multiplies when workers spawn additional sub-workers beyond the first level
When does Lesson 28 recommend using a tool instead of a dedicated worker agent?
When the subtask is a single model call with a well-defined schemaWhen the subtask needs multiple steps and its own memory across those stepsWhen the subtask is slow and you want to run it in parallel for speedWhen independent deployability of the subtask component is required