The Infrastructure That Makes Agent Networks Possible
The Infrastructure That Makes Agent Networks Possible
Everyone focuses on how agents communicate. The real bottleneck is shared state.
Communication Is Solved
Agents can send messages to each other. They can use APIs, message queues, shared files, or direct function calls. The mechanics of agent-to-agent communication are a solved problem. What is not solved is agents sharing a consistent view of the world.
The Shared State Problem
When two agents work on the same codebase, they need to see the same files. When three agents manage a customer pipeline, they need to see the same CRM state. When five agents handle different parts of a deployment, they need to see the same infrastructure state.
Without shared state, every agent operates on its own snapshot of reality. These snapshots diverge immediately. Agent A reads a file, Agent B modifies it, Agent A acts on the stale version. The result is conflicts, overwrites, and corrupted state.
What Good Infrastructure Looks Like
The infrastructure that makes agent networks work:
- Git as shared state for code - branches, worktrees, and merge resolution give agents a shared, versioned filesystem
- Database as shared state for data - transactional reads and writes prevent conflicts
- File locks for exclusive access - when an agent needs to modify something, it claims it first
- Event streams for notifications - agents subscribe to changes rather than polling
The pattern is not new. This is the same infrastructure distributed systems have used for decades. The difference is that the "distributed nodes" are now AI agents instead of microservices.
The Implication
Building agent networks requires distributed systems expertise, not AI expertise. The hard problems are consistency, conflict resolution, and state management - all problems with known solutions that most agent framework builders are ignoring.
Fazm is an open source macOS AI agent. Open source on GitHub.