Multi-agent costs multiply
A manager calls three specialists, each specialist uses five tool loops, the reviewer runs twice, and suddenly one user request became dozens of model calls. That may be worth it. It is not free.
Debugging also multiplies. You need traces per agent, shared correlation ids, handoff summaries, and a way to replay only the failed branch.
Budget before architecture
Set per-agent step limits and total run limits. Track cost by branch, not only by whole run. If a worker has no useful result after N steps, stop it and return partial failure.
Parallelism improves wall-clock time but can increase total cost. Measure both.