OpenAI Codex developer Eric Provancher warned that launching more than two parallel AI sub-agents almost always increases token consumption without noticeable quality improvement. According to his description, agents re-verify each other's work due to a lack of trust; he termed this effect the "coordination tax."
As an example, Provancher cited a Python refactoring project: 1 393 agents spent tokens costing 20 000 dollars on it, although, in his estimation, a single Astra agent could have completed the task for a significantly lower amount.
The practical implication of the statement is not a complete rejection of parallel work, but rather the necessity of comparing its benefits against the costs of reconciling results. Meanwhile, the source materials do not allow for verification of project details, result quality, or the exact magnitude of savings.
editorial commentary
Why it matters
A likely consequence is a stricter evaluation of the return on investment for parallel AI systems per task. The next observable signal will be the publication of primary data on the quality, time, and costs of such runs. Significant uncertainty remains because currently only a metadata-only synopsis from one source is available.