Cross-Machine AI Orchestrator
Autonomous task dispatch across 3 machines
Complex software projects require work that spans multiple environments: a planning agent that designs architecture, executor agents that write and test code, and a coordinator that manages dependencies between tasks. Running all of this on a single machine creates resource contention, and there was no existing framework for coordinating AI agents across physical machines with persistent state.
I built a 3-tier orchestration framework: a controller machine dispatches tasks to two worker machines via SSH tunnels and SQLite message queues. The controller maintains a DAG (Directed Acyclic Graph) of task dependencies, ensuring work is executed in the correct order.
Each worker runs autonomously, processing tasks from its queue, executing code changes, running tests, and reporting results back. The system handles error recovery with configurable retry policies, state persistence across restarts, and convergence detection for iterative refinement loops.
A generative refinement loop allows the planner to evaluate worker output, identify issues, and dispatch corrective follow-up tasks automatically. The system tracks generation history and detects oscillation or plateau patterns to know when to stop iterating.