A lot of multi-agent systems are an org chart drawn in software: a researcher agent, a writer agent, a critic. That shape adds a coordination tax without adding capability when the sub-tasks were never independent. Multi-agent earns its cost when four conditions hold at once, and four questions check them. List the jobs you are thinking of handing to a fleet, answer the four on each, and get a ranked call: one agent, a hybrid or a team, with the answers that decided it.
Built for the essay Your Multi-Agent System Is an Org Chart
Add the jobs you are considering splitting across agents, including one you have already built. Pick the closest shapes or type your own. Judge the real task in front of you, with its real data and edge cases.
Ranked by whether multi-agent earns its coordination tax. One agent means the parts depend on each other or write one result (or you could not tell), so a split would be one job in a team costume. Tap any task to see what decided it, its architecture, its cost and the anti-pattern to avoid.
Order: a team first, then a hybrid, then one agent. Within each, a task that meets more of the four conditions ranks higher; ties keep the order you added them.
Token figures: Anthropic, How we built our multi-agent research system (13 June 2025). An agent used about 4× the tokens of a chat; their multi-agent system about 15×. Checked at the source on 7 Oct 2026.
Free. New tools and the essays behind them, with every source shown. The button opens Substack; nothing you entered here goes with it.
Subscribe free