The AI Swarm Hype is Dead: Why Multi-Agent Systems are Burning Your Budget

We’ve all seen the flashy GitHub repos and VC pitch decks promising a future where a "swarm" of AI agents magically builds your next startup while you sip margaritas. It sounds amazing on paper. You have a manager agent, a developer agent, a QA agent, and maybe a designer agent all chatting in a digital Slack room. But if you've actually tried running these setups, you probably noticed two things: they take forever, and your API bill looks like a phone number.

Now, OpenAI Codex developer Eric Provencher has put some hard numbers on what many of us suspected. He warns that running more than two parallel sub-agents is essentially burning money with zero quality gain. Provencher points to a mind-boggling project where a swarm of 1,393 agents spent a whopping $20,000 in tokens to refactor a single Python codebase—a task that a single advanced reasoning model like GPT-6 Astra could have probably handled for pennies on the dollar.

Provencher calls this the "coordination tax." In short, agents don't trust each other. When you put a bunch of LLMs in a loop, they spend most of their time double-checking, correcting, and arguing over each other’s work. It’s the digital equivalent of a bloated corporate committee meeting where nothing gets done, but everyone gets paid by the hour.

As someone who tests these tools daily, I’ve felt this pain firsthand. The UX of swarm frameworks is admittedly mesmerizing—watching terminal windows scroll with "Agent A is thinking..." feels like living in the future. But the utility just isn't there yet. Most multi-agent frameworks are bloated, unoptimized, and frankly, lazy engineering. Developers are throwing more agents at a problem instead of designing better prompts, building solid deterministic guardrails, or using single-agent reasoning loops.

For the AI ecosystem, this is a much-needed reality check. The era of "agentic bloat" is hitting a wall of economic reality. If you are building AI workflows today, skip the 10-agent swarm. Stick to one high-quality reasoning model, give it a solid system prompt, and maybe—just maybe—pair it with a single critic agent to review the output. Your wallet will thank you.

Photo: Tyler / Unsplash (https://unsplash.com/@tylergm)