Multi-agent systems split work across several specialised agents.
A common shape is an orchestrator delegating to focused sub-agents.
Each sub-agent gets a narrower context and a smaller tool set.
Coordination overhead and error propagation are the real costs.
Try a single well-scoped agent first; add agents only when one clearly cannot cope.
A multi-agent system splits work across several agents with different roles.
The common shape is an orchestrator delegating to focused sub-agents.
A researcher, a writer and a reviewer is the standard illustration.
The appeal is that each gets a narrow context and a small tool set.
Narrow context genuinely helps, since a smaller prompt with fewer tools performs better.
The costs arrive quickly and are usually underestimated.
Every hand-off is a summary, and every summary loses information.
Errors propagate, since a sub-agent's wrong answer is accepted as fact by the next.
Debugging is much harder, because failure may be in any agent or in a hand-off.
Cost multiplies, as five agents means five conversations rather than one.
Latency multiplies too when agents run in sequence.
Agents that talk to each other can loop, agreeing politely and achieving nothing.
Try one well-scoped agent first and let a genuine limit justify more.
The clearest wins are parallel independent subtasks, such as researching ten companies at once.
There the split is real work divided, not a conversation multiplied.
One question on this card. Every option explains itself once you pick, right or wrong.
17 concepts in this module, then a graded test with every answer explained.