When should you use a single agent versus a multi-agent...

The default should be a single agent, and multi-agent should be a decision you’re forced into by a specific limitation rather than a starting architecture. The legitimate reasons to add agents:

• Parallelism — independent sub-tasks that can genuinely run at the same time. Analyzing 50 markets is 50 independent analyses; a single agent doing them sequentially takes 50x as long as agents doing them in parallel.
• Specialization — sub-tasks needing different tools, different system prompts, or even different models perform better with an agent tuned for that one job than one generalist trying to hold all of it.
• Context overflow — the task is too large for one agent’s working context, so it has to be split and processed in pieces regardless of parallelism.
• Quality through separation — one agent generates, a separate one critiques; the separation itself improves output over one agent doing both roles in the same pass.

If none of those apply — the task is single-domain, sequential, and fits comfortably in context — a multi-agent system just adds coordination overhead, more places for context to get lost at a handoff, and more latency for no measurable gain. Reach for the split because you hit one of the four walls above, not because it mirrors how a human team would be organized.