OpenAI Codex developer warns against oversized AI agent swarms
Eric Provencher, a developer working on OpenAI Codex, has criticized the use of large groups of parallel AI agents in posts on X. He says that running more than two sub-agents at once usually consumes additional tokens without producing better results. In his view, the agents often lack trust in one another and repeatedly check work that has already been completed.
Provencher describes this overhead as a “coordination tax” that makes it difficult to manage many parallel workstreams without excessive verification and token use. He cited a project in which 1,393 Fable agents were used to refactor a single Python file at a reported cost of $20,000 in tokens. According to Provencher, one Astra agent could have completed the task for a fraction of that cost.
As a possible alternative, he suggests assigning work to separate threads that contact the main agent only after finishing, rather than being polled continuously. He also noted that system prompts are executed repeatedly across sub-agents and that missing context can lead to duplicated work. Provencher said OpenAI still needs to develop better solutions.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.