How Shifting from Orchestrators to Supervisors Scaled an AI Agent Fleet

A development team rebuilt their AI automation system by moving from orchestrator-based pipelines to a supervisor model, where agents receive assignments rather than step-by-step instructions. The original fleet used scheduled jobs and small models for tasks like sorting mail and bundling reports, but adding new capabilities meant adding entire new workflows, making the system harder to maintain. The breakthrough came when agents were given the ability to act autonomously — editing files, running tests, and connecting to external tools via the MCP open standard — all operating within Claude Code. Under the new supervisor model, agents are handed a scoped assignment with explicit goals, non-goals, permitted tools, and measurable acceptance criteria instead of a pre-planned sequence of steps. The team notes that this shift places greater responsibility on writing clear assignments upfront, since no intermediary exists to fill gaps during execution.