Eric Provencher, a developer at OpenAI Codex, has pointed out that deploying more than two parallel AI sub-agents often leads to wasted tokens without any improvement in output quality. According to The Decoder, Provencher explained that this inefficiency arises because agents tend to distrust each other and engage in redundant double-checking.
The Decoder also reported on a project involving 1,393 agents that spent $20,000 worth of tokens on a single Python refactoring task. Provencher noted that the same task could have been completed much more efficiently by a single Astra agent, dramatically reducing costs.
For Japanese investors and tech firms exploring AI applications in FX, crypto, and equities trading, Provencher’s insights underscore the importance of optimizing AI resource allocation to control operational expenses and improve system performance.