Aggregate vs. Per-Sub-Agent Budgets

Your deep research pipeline caps each sub-agent at 8 tool-call steps and a 3-minute timeout, individually. Five sub-agents are dispatched in parallel for a query.

1. Explain a concrete scenario where every sub-agent stays within its individual budget, yet the pipeline still fails to hit its overall 5-20 minute latency target.
2. What additional control does this scenario argue for, and how does it relate to (not replace) the per-sub-agent budgets already in place?
3. Describe the correct graceful-degradation behavior when the aggregate budget is hit while 2 of 5 sub-agents are still running, and contrast it with two wrong ways to handle it.