The article has been written to `/var/www/simpleprog-website/papers/avatar-llm-workflow-orchestration.html` (65 lines).
Here's a summary of what the article covers:
**Opening** - Frames the problem: scientific workflow systems use rigid, hand-tuned orchestration rules that cannot adapt to runtime conditions, and retrofitting LLM agents into existing systems is architecturally messy.
**Three-Actor Architecture** - Avatar decomposes workflow orchestration into three actors (orchestrator, executor, provenance monitor) communicating through a shared message fabric. Each actor's decision policy is pluggable: rule-based or LLM-backed, validated through adapters against a fixed action catalog.
**Implementation** - Built on Academy middleware, with concrete actors for provenance diagnosis, task execution, and workflow control. Three modes: M0 (all rules), M1 (LLM diagnosis only), M2 (LLM diagnoses and decides).
**Evaluation** - Three workloads: resilience (fault-prone DAG), scaling (bursty streaming pipeline), and active learning (molecular design campaign). The same unchanged core runs all three.
**Key Finding** - LLM reasoning is not universally better. It excels at infrequent, context-dependent decisions (diagnosing permanent failures, detecting convergence) but loses to simple rules on high-frequency control (autoscaling) due to reasoning latency.
**Practical Takeaway** - Avatar achieves 55% reduction in compute wastage and 40% cut in GPU-busy time by applying LLM reasoning at specific leverage points, not wholesale replacement of rule-based orchestration.