/classifier-lab-subagent

Validate whether subagents can reliably run long-running deterministic tasks.

The Experiment

3 subagents run concurrently, each training a different text classifier model on the same HuggingFace dataset (ag_news). Each subagent runs /classifier-lab's existing self-improvement loop. The orchestrator watches JSON streams, enforces timeouts, and collects results.

The point: Determine if subagents can reliably perform long-running tasks when 90% of the intelligence is in the code and the subagent is just a process runner.

Architecture