Two multi-agent designs are compared for a research task. In one, a planner, a coder, and a reviewer agent all belong to the same conversation and exchange messages directly with each other in turn, with each agent seeing what the others said. In the other, a single lead agent breaks the task into independent subtasks and spawns several subagents that each work on their own subtask in parallel, but the subagents never see each other's work or communicate with one another -- each one reports its findings back only to the lead agent, which synthesizes everything itself. What is the defining structural difference between these two designs?
- The first design cannot use more than one tool across the whole conversation, while the second allows each subagent its own tools
- The two designs are functionally identical, since both ultimately involve more than one agent contributing to a single overall task
- The first design is peer-to-peer: every participating agent can see and respond to every other agent's messages within one shared conversation. The second is strictly hierarchical: subagents work independently and in isolation from each other, with all coordination flowing one level up through the lead agent, trading away emergent cross-agent problem-solving for simpler, more predictable coordination
- The second design requires that every subagent be given an identical prompt and role, since giving them different instructions would make the lead agent unable to synthesize their reports
Why C? And why not the others?
Correct answer: C. The first design is peer-to-peer: every participating agent can see and respond to every other agent's messages within one shared conversation. The second is strictly hierarchical: subagents work independently and in isolation from each other, with all coordination flowing one level up through the lead agent, trading away emergent cross-agent problem-solving for simpler, more predictable coordination
A group-chat-style design lets every participating agent read and respond to what every other agent in the conversation says, so a planner, coder, and reviewer can react to each other's messages directly within one shared thread; an orchestrator/lead-and-subagents design instead keeps subagents isolated from each other entirely -- each independently works its own piece of the task and reports back only to the lead agent, which is the sole point where results are combined -- trading the possibility of subagents catching or building on each other's work directly for a simpler, more predictable, strictly hierarchical coordination structure. The option limiting the first design to a single tool across the whole conversation is wrong: nothing about agents conversing directly with each other restricts how many tools are available across the group; each participant can still have its own tools. The option calling the two designs functionally identical is wrong because it erases the exact distinction being tested -- peer-to-peer visibility versus strict hierarchy with no lateral communication. The option requiring identical prompts and roles for every subagent is wrong: subagents in the hierarchical design can be given distinct subtasks and instructions, and the lead agent's synthesis step works specifically because each one reports back on its own distinct piece of the work.
Source: Anthropic, 'How we built our multi-agent research system' (anthropic.com/engineering/multi-agent-research-system) -- describes an orchestrator-worker pattern where subagents operate in parallel, do not communicate with each other, and report findings back to a LeadResearcher which synthesizes results