Dumb Orchestrator With Smart Workers Beats One Big Agent
Dumb Orchestrator With Smart Workers Beats One Big Agent
The instinct when building an AI agent system is to make one powerful agent that handles everything. Feed it all the tools, give it a big context window, and let it figure things out. This works for demos. It falls apart in production.
The Single Agent Problem
A monolithic agent that handles browser automation, file management, accessibility tree navigation, and sequential multi-step workflows simultaneously has too many failure modes. When it gets confused about which tool to use, the entire pipeline breaks. Error recovery is a mess because the agent's context is polluted with state from every subsystem.
The Orchestrator Pattern
The approach that actually works: a dumb manager with good routing rules, plus specialized workers. The orchestrator does not need to be smart. It needs a decision tree.
- Does the task involve a web browser? Route to the browser agent.
- Does it need to interact with a native macOS app? Route to the accessibility agent.
- Is it a multi-step workflow with dependencies? Route to the sequential agent.
Each worker agent is scoped to one domain. The browser agent only knows about Playwright, page navigation, and web element interaction. The accessibility agent only knows about the macOS accessibility tree, AXUIElement, and native app controls. They are experts in their lane.
Why This Is More Reliable
When a specialized worker fails, the orchestrator knows exactly what failed and can retry or fall back. The worker's context is clean - it is not confused by irrelevant tool definitions or state from other subsystems.
The routing logic itself is simple enough to be deterministic. You do not need an LLM to decide whether a task involves a browser - pattern matching on the task description handles 95% of cases.
The 3-Hour to 4-Minute Story
One team applied this pattern to their content pipeline - what used to take 3 hours of manual work collapsed to 4 minutes. The key was not making a smarter agent. It was making simpler agents with clear boundaries and letting a basic router connect them.
Fazm is an open source macOS AI agent. Open source on GitHub.