How Is Everyone Creating Multiple Agents Under One Orchestrator
How Is Everyone Creating Multiple Agents Under One Orchestrator
The multi-agent orchestration question comes up constantly. You want a coding agent, a research agent, and a communication agent working together. The naive approach is one big prompt with multiple personalities. That breaks immediately.
The Soul File Approach
A soul file is a persistent configuration that defines an agent's identity, capabilities, and boundaries. Instead of one orchestrator with sub-prompts, each sub-agent gets its own soul file. The orchestrator's job is routing, not role-playing.
Each soul file contains: what this agent does, what tools it has access to, what it should never do, and how it communicates results back. The orchestrator reads incoming tasks, matches them to the right sub-agent, and dispatches. It does not try to be all agents at once.
Clear Scope Prevents Chaos
The biggest mistake in multi-agent systems is overlapping scope. Two agents both think they handle email. Three agents all have file system access with no coordination. The result is conflicts, duplicate work, and corrupted state.
Clear scope means each agent owns a domain exclusively. The email agent handles all email. The coding agent handles all code changes. The research agent handles all web lookups. When a task crosses domains - like researching a bug and then fixing it - the orchestrator chains the agents sequentially, passing context between them.
Persistent Memory per Agent
Each sub-agent maintains its own memory. The coding agent remembers your codebase conventions. The email agent remembers your communication style. This memory persists across sessions through the soul file and its associated state.
The orchestrator does not need to remember everything. It needs to remember which agent knows what. This keeps each context window focused and prevents the context overflow that kills monolithic agent setups.
Fazm is an open source macOS AI agent. Open source on GitHub.