Agents
Each major LoopAI capability can be understood as a composable agent or subgraph module.
StarterAgent
Starter is the coordinator. It mainly handles:
- user interaction
- intent detection
- choosing the right execution path
- chaining together downstream agent work
If LoopAI is treated like an operating system, Starter is closest to the task scheduler.
JudgerAgent
Judger focuses on current model quality. It usually handles:
- running evaluations
- comparing results
- locating failed samples
- providing evidence for later analysis
When local or remote OpenAI-compatible services are configured, Judger can work across different inference backends.
AnalyzerAgent
Analyzer turns observations into conclusions:
- grouping failure patterns
- analyzing likely causes
- generating actionable optimization suggestions
It connects evaluation results with data strategy and is a key part of the loop.
TrainerAgent
Trainer carries out the actual optimization step, such as:
- invoking training frameworks
- launching asynchronous jobs
- collecting training logs
- sending results back into system state
Local training is typically integrated with LLaMA-Factory or verl.
Why split the system into agents
This split gives the system three clear advantages:
- Each module has a narrower responsibility and is easier to replace or test.
- New capabilities can be added without rewriting the entire flow.
- Teams can choose which steps to automate and which steps to review manually.