OpenAI's Provencher: >2 sub-agents burn tokens, zero quality gain

Eric Provencher's warning lands as AI agent swarms — setups where many parallel sub-agents tackle one task — are being marketed as a way to speed up work. His X posts describe the failure mode in concrete terms: agents don't trust each other, so they double-check everyone's homework, and the token bill climbs while output quality stays flat. The example he cites is stark: 1,393 Fable agents spent $20,000 in tokens refactoring a single Python file, work he says one Astra agent could have done for a fraction of the cost.

The mechanics he points to are mundane but costly. System prompts add up across sub-agents, and without enough context they make redundant tool calls; polling for status keeps the main agent busy instead of waiting for a notification. He suggests delegating to separate threads that notify the main agent only once done, and acknowledges OpenAI still needs to ship better solutions here. Swarms may save time, he writes, but the token overhead is a "trap."

The stakes hinge on whether better orchestration — smarter context sharing, notification-based handoffs — can cut the coordination tax without losing the speed advantage. For teams already running multi-agent pipelines, the reading is that parallel sub-agents above two are likely a cost center rather than a quality lever, at least until the tooling catches up.