A Language for Describing Agentic LLM Contexts

Abstract

Large language models are increasingly used within larger systems ("LLM agents"). These systems often make a sequence of LLM calls, where each call provides the LLM with a combination of instructions, observations and interaction history, among other information. The design of the encoded information and its structure play a central role in the quality of the resulting system, as evidenced by the term "context engineering". Despite this, it remains remarkably difficult to communicate the composition of the LLM context in a system, and how it evolves over time. While verbalizing the language of an individual prompt is standard and prompt templates are often included in a system's description, no such system exists for communicating the structure and content of an evolving context. Whether between teammates iterating on an agent's behavior or researchers describing a working system in a publication, context construction is typically conveyed through hand-waving, informal prose, ad hoc diagrams, or direct inspection of code---none of which precisely capture how a prompt evolves across interaction steps or how two context representation strategies differ from each other. To remedy this, we introduce the Agentic Context Description Language (ACDL), a visual language for specifying the structure and dynamics of LLM input contexts in a precise, readable and standard manner. ACDL provides constructs for specifying context aspects such as role message sequences, dynamic content, time-indexed references, and conditional or iterative structure, capturing the full architecture of a prompt independently of any particular implementation. ACDL diagrams can be hand drawn on a whiteboard, or written in formal language which can then be rendered. We describe the language, demonstrate it by documenting several existing systems and their variants, and encourage the community to adopt it for describing LLM systems, both in day-to-day communication and in papers.