OpenAI Codex dev warns AI agent swarms waste tokens

eric provencher, a developer at openai codex, argues that ai agent swarms are inefficient and waste tokens. he claims that more than two parallel sub-agents almost always burn tokens without improving quality. provencher attributes this to the 'coordination tax' and redundant work caused by sub-agents lacking context. for example, a project spent $20,000 in tokens refactoring a single python file with 1,393 fable agents, which could have been done by a single astra agent for a fraction of the cost. this inefficiency affects ai project costs and efficiency.