Adding more AI agents to a coding task can raise the bill without improving the result, according to OpenAI Codex developer Eric Provencher. He says systems with more than two parallel sub-agents often spend tokens verifying one another’s work and repeating the same tool calls.
Provencher calls the effect a “coordination tax.” Each sub-agent consumes its own system prompt and needs enough context to act usefully. When that context is incomplete, agents repeat research or inspect work that another agent has already checked. Constant status polling adds another layer of traffic without advancing the task.
He pointed to an experiment that reportedly used 1,393 agents and $20,000 worth of tokens to refactor one Python file. In his view, a single capable model could have completed that job for a fraction of the cost. The example is anecdotal rather than a controlled benchmark, but it illustrates how parallelism can optimize elapsed time while making total compute dramatically worse.
One suggested pattern is to give independent tasks to separate threads and have each notify the main agent only when finished. The broader lesson for developers is to measure completed work per token, not agent count. Parallel workers help only when tasks are genuinely separable and their results do not require expensive mutual verification.