OpenAI’s former Codex developer, Eric Provencher, has raised concerns about the inefficiency of AI agent swarms, claiming they are a significant waste of computational resources with no corresponding gain in quality. According to Provencher, when more than two AI agents work in parallel, they often end up duplicating efforts due to a lack of trust and coordination, leading to what he terms the 'coordination tax.'
The Cost of Coordination
Provencher highlighted a specific project in which 1,393 agents were deployed to perform a single Python refactoring task. The endeavor consumed $20,000 in tokens—equivalent to the cost of having one expert agent complete the same task for a fraction of the price. This example underscores a critical issue in AI development: the increasing complexity of multi-agent systems often does not translate into proportional improvements in output quality.
Implications for AI Development
This critique comes at a time when AI agent swarms are being increasingly adopted in enterprise and research environments, especially in tasks requiring collaborative problem-solving. Provencher’s findings suggest that developers should carefully evaluate the trade-offs between parallelization and efficiency. The high token cost and lack of quality improvement may deter organizations from scaling up AI systems in this manner, pushing them toward more streamlined, single-agent solutions.
Looking Ahead
As AI systems become more sophisticated, the challenge of optimizing agent collaboration will remain central to effective deployment. Provencher’s observations serve as a timely reminder that innovation in AI doesn’t always mean adding more components—it may mean smarter coordination or simpler, more focused tools.