Multi-Agent Hype vs Economic Reality in Production
Multi-Agent Hype vs Economic Reality in Production
Multi-agent architectures are having a moment. The pitch sounds great: a planner agent breaks down the task, an executor agent does the work, a reviewer agent checks the output. Clean separation of concerns. Beautiful in theory.
In practice, someone tried this and found it was burning roughly 3x the tokens of just running one good agent with a solid prompt. That is not a rounding error - it is the difference between a sustainable product and one that bleeds money.
Where the Tokens Go
Each agent in the chain needs context. The planner reads the task description, reasons about it, and produces a plan. The executor reads the plan plus the original context and does the work. The reviewer reads the original task, the plan, and the output, then produces feedback.
Every handoff duplicates context. Every agent re-reads overlapping information. And when the reviewer sends feedback back to the executor for a revision, you get another full cycle of token consumption.
When Multi-Agent Actually Pays Off
Multi-agent makes economic sense in specific scenarios:
- Genuinely parallel work where agents tackle independent subtasks simultaneously
- Different model tiers where a cheap model plans and an expensive model executes only the hard parts
- Compliance requirements where separation between execution and review is mandated
- Very long tasks where a single context window cannot hold everything
The Single-Agent Alternative
A well-structured single agent with a detailed system prompt, clear CLAUDE.md instructions, and good tool access often outperforms a multi-agent chain. It has full context at all times, no handoff overhead, and no communication failures between agents.
The key insight is that adding agents adds communication overhead. Unless that overhead buys you something concrete - parallelism, model routing, compliance - you are paying extra for complexity that makes the system harder to debug and more expensive to run.
The Bottom Line
Start with one agent. Make it great. Only add agents when you have a specific problem that a single agent cannot solve. The economics of LLM inference mean that every unnecessary agent conversation directly hits your margin.
Fazm is an open source macOS AI agent. Open source on GitHub.