Eric Provencher, a developer behind OpenAI’s Codex, is pulling the emergency brake on the current hype around massive AI agent swarms. In a series of candid posts on X, he argues that deploying more than two parallel sub-agents rarely improves output quality. Instead, it triggers what he calls a “coordination tax.” The agents don’t trust each other’s work, leading to excessive verification loops where they essentially double-check everyone’s homework. This redundancy burns through tokens at an alarming rate without delivering better code or insights.
The financial implications are stark. Provencher highlighted a specific case where a user spent $20,000 in token costs just to refactor a single Python file. The culprit? A swarm of 1,393 Fable agents working in parallel. According to Provencher, a single Astra agent could have handled the same task for a fraction of that cost. While proponents argue that swarms save human time, Provencher labels the token overhead a “trap.” The sheer volume of redundant computation offers diminishing returns, turning what should be an efficiency tool into a budgetary black hole.
The root cause lies in how these systems communicate. Sub-agents often lack full context, forcing them to request information repeatedly. System prompts get executed multiple times across different threads, compounding the waste. Provencher suggests a structural fix: delegate tasks to separate threads that only notify the main agent upon completion, rather than constantly polling for status updates. This reduces the chatter and keeps the workflow lean. He admits, however, that OpenAI still needs to ship better native solutions to manage this complexity efficiently.
This critique exposes a growing disconnect between academic benchmarks and real-world economics. Many developers are chasing the aesthetic of complex multi-agent architectures without auditing the unit economics. If a $20,000 refactoring job can be done for pennies by a single model, the industry needs to rethink its obsession with scale. Complexity isn’t always sophistication; sometimes it’s just expensive noise. Until platform-level optimizations arrive, developers should treat agent swarms with extreme skepticism.
