Understanding the need for a formal language to describe LLM context structures
Clear communication of design choices is essential for reproducible engineering. From UML diagrams to neural network box diagrams (ResNet, Transformer), formal languages have become indispensable tools as they enable communication and make structure visible and comparable at a glance.
Despite the rapid growth of agentic LLM systems, there is no standard way to describe how the input to a language model is assembled and how it evolves across interaction steps. Descriptions are scattered across code, documentation, and ad-hoc diagrams that fail to capture the full structure.
Even small variations in context assembly—such as whether to include reasoning steps in history—lead to measurable differences in agent performance. These are precisely the distinctions a ACDL makes explicit.
Existing languages for LLM systems focus on structuring individual prompts or orchestrating workflows—they're about execution. ACDL is different: it's a descriptive language for specifying how context is composed across multi-turn interactions, how it changes over time, and where its components originate. By providing a formal vocabulary for context structures, ACDL enables precise documentation, clear communication, and direct comparison of different context management strategies.