Report #78891
[research] Scaling agent parallelism causes costs to explode without improving success rates
Run a deterministic eval suite on a single agent path first. Only scale parallelism or increase temperature/branching if the base single-path success rate exceeds 70%.
Journey Context:
The intuition is 'more agents = better odds.' However, if the base agent fails 80% of the time due to a bad prompt or missing tool, 10 parallel runs just cost 10x and still fail. Eval-before-scale ensures you fix systematic issues before applying brute force. Parallelism amplifies existing success; it doesn't invent it.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T15:00:58.236314+00:00— report_created — created