mlld:orchestrator

You build resilient LLM workflows by designing mlld orchestrators that coordinate pipelines and enforce crash-resilience. These orchestrators apply checkpoint-based resumption to automatically recover from interruptions and resume work.

The dumb orchestrator pattern coordinates end-to-end LLM calls by relying on prompts to drive decisions and structured JSON action schemas. The orchestrator manages the loop and parallel fan-out while delegating decisions to the prompts.

Parallel worker calls are set up within mlld orchestrators by applying parallel fan-out across LLM workflows. This distributes calls simultaneously, with caching and validation enforcing resilience across the distributed tasks.

Yes, mlld orchestrators support ad-hoc decision loops tailored for development and research archetypes. You can implement decision-agent driven phase inference to dynamically adjust the workflow path based on intermediate LLM outputs.

Structured JSON action schemas are validated within mlld orchestrators by applying validated invalidation to LLM outputs. This enforces structured prompts and JSON schemas, ensuring the workflow only advances when actions meet defined requirements.

The best way to start is by initializing a new orchestrator with mlld init. You then implement a minimal llm/run workflow to validate the core loop before scaling to multi-phase pipelines and parallel fan-out.