Eric Provencher, a Codex developer at OpenAI, warned in public posts on X that using large parallel agent swarms is an inefficient use of tokens that fails to yield quality improvements. Provencher noted that running more than two parallel sub-agents creates a costly “coordination tax,” as agents distrust one another and waste compute resources verifying peer output.
According to Provencher, sub-agents frequently duplicate effort due to system prompt overhead and a lack of sufficient context, leading to redundant tool calls. Highlighting an extreme case where $20,000 in tokens was spent using 1,393 Fable agents to refactor a single Python file, he asserted that a single agent could have executed the task for a fraction of the cost.
To mitigate excessive token burn, Provencher suggested delegating sub-tasks to isolated threads that notify the primary agent only upon completion, rather than continuously polling for status updates. He acknowledged that OpenAI still needs to ship improved native solutions to address these agent orchestration challenges.
Why it matters
Cap parallel sub-agents at two per workflow to avoid exponential token burn and redundant tool calls.
Implement isolated execution threads with completion notifications instead of continuous status polling to reduce system prompt context bloat.
Evaluate single-agent architectures before deploying multi-agent swarms, which often introduce costly coordination taxes without measurable quality gains.
Source: the-decoder.com
