Simple Routing Rules Beat Complex Orchestrators for Parallel AI Agents
Simple Routing Rules Beat Complex Orchestrators for Parallel Agents
The delegation routing problem is one of the hardest parts of running multiple AI agents in parallel. You have several specialized agents, each with different capabilities, and you need to decide which agent handles which task. Most people overcomplicate this.
Why Complex Orchestration Fails
The instinct is to build a smart orchestrator - an LLM that reads incoming tasks, evaluates agent capabilities, and routes intelligently. The problem is that this orchestrator itself consumes tokens, adds latency, and introduces a single point of failure. When it misroutes a task, debugging gets painful fast.
What Actually Works
Simple routing rules based on file paths and task types. If the task touches UI files, it goes to the frontend agent. If it involves API endpoints, it goes to the backend agent. If it's tests, it goes to the test agent.
These rules are dumb and rigid and they work better than anything clever. Each agent gets a clear boundary, a clear set of files it owns, and clear instructions not to touch anything outside its scope.
The Codebase Ownership Model
Think of it like team code ownership in a large engineering org. Each agent "owns" a directory or module. Ownership means the agent reads and writes only within its boundaries. Cross-boundary work gets queued for the appropriate agent rather than handled inline.
This eliminates most merge conflicts, reduces wasted tokens from agents reading irrelevant files, and makes the whole system predictable.
Voice I/O Adds Another Layer
When you add voice input and output on top - like running agents on a Raspberry Pi or a local Mac - the routing becomes even more important. Voice commands tend to be ambiguous, so having simple, predictable routing rules means you can map spoken intent to the right agent without needing another LLM layer to interpret.
Keep it simple. Route by file path. Own your boundaries.
Fazm is an open source macOS AI agent. Open source on GitHub.