Running more than two parallel sub-agents almost always burns tokens without improving output quality, warns Eric Provencher, a Codex developer at OpenAI. Massive multi-agent setups trigger what Provencher terms a coordination tax, where autonomous workers waste compute verifying each other rather than advancing the task.
The scale of this inefficiency surfaced in a project where an engineer spent $20,000 in tokens refactoring a single Python file using 1,393 agents. Provencher noted that a single agent could have completed the refactoring for a fraction of the cost. While swarms may save time, Provencher writes that the underlying token overhead is a financial trap for engineering teams.
The mechanism behind the waste is straightforward: system prompts accumulate across sub-agents, leading to redundant tool calls due to inadequate context isolation. Provencher notes that duplicate effort remains a structural flaw in agentic workflows. Instead of continuous polling, he suggests delegating tasks to separate threads that notify the main agent only upon completion, while acknowledging that OpenAI still needs to ship better orchestration layers for these deployments.
Scaling agent counts is an expensive substitute for proper execution architecture.