Individuals Get Smarter with LLMs, Groups Get Dumber
Individuals Get Smarter with LLMs, Groups Get Dumber
A single Claude Code agent is remarkably productive. It understands context, writes good code, and completes tasks faster than most developers. Add a second agent and you get nearly double the throughput. Add a third and you get maybe 2.5x. By the fifth agent, you are back to 3x because they spend as much time dealing with each other's changes as they spend on actual work.
The Coordination Tax
Every additional agent adds a coordination cost that grows non-linearly. With 2 agents, there is 1 potential conflict pair. With 5 agents, there are 10. With 10 agents, there are 45. Each conflict pair has a chance of producing a file collision, a build conflict, or a logic inconsistency.
The individual agent does not get dumber - each one is still doing brilliant work on its assigned task. But the system as a whole produces worse results because the brilliant individual outputs conflict with each other.
Why This Mirrors Human Teams
Brooks's Law from 1975 - "adding manpower to a late software project makes it later" - applies directly to AI agents. The communication overhead eventually exceeds the productivity gain. AI agents actually hit this wall faster than humans because:
- They lack implicit social coordination
- They cannot informally check with each other
- They take instructions literally without reading between the lines
- They do not naturally back off when they sense a conflict
The Sweet Spot
For most codebases, the optimal number of parallel agents is:
- 1-2 for small projects (< 10k lines)
- 3-4 for medium projects with clear module boundaries
- 5-6 for large projects with strict file ownership rules
- Never more than 8 unless you have a dedicated coordination system
Making Groups Smarter
The fix is not smarter individual agents - it is better coordination protocols. File locking, build queues, explicit ownership declarations, and human-managed merge points. Treat multi-agent development like distributed systems engineering, because that is exactly what it is.
Fazm is an open source macOS AI agent. Open source on GitHub.