Recoleta
GitHub
Source note

A Language for Describing Agentic LLM Contexts

Authors
Noga Peleg Pelc, Gal A. Kaminka, Yoav Goldberg
Source
arXiv
Published
Collection
Software Intelligence
LLM AgentsContext EngineeringPrompt ArchitectureMulti Agent SystemsAgent Documentation
Open arXivMarkdown source

Summary

ACDL is a descriptive language for specifying how LLM agent contexts are assembled and how they change over time. It targets clearer comparison, reproduction, and team communication for agent systems.

Problem

  • Agent papers and code often describe context construction with prose, ad hoc diagrams, or implementation details, which leaves key prompt-history behavior ambiguous.
  • This matters because context layout affects agent behavior, including which messages, tool outputs, reasoning traces, and prior turns the LLM sees at each step.
  • Reproducing or comparing agent systems is hard when papers do not state how state and history are mapped into LLM inputs.

Approach

  • ACDL describes the context window as a sequence of role messages and information pieces, using symbolic labels instead of exact prompt wording.
  • It tracks time-indexed state such as env.user_input[@T], system state such as sys.conf.role, and prior model responses through resp references.
  • It supports conditions, loops, named expressions, fragments, nested time steps, and multi-agent contexts.
  • The language stays descriptive: it specifies what the LLM receives, not how tools, retrieval, memory, or agent control logic work.
  • The paper also provides visual diagrams, a parser, an interactive renderer, a VS Code plugin, examples, and an agentic skill.

Results

  • The excerpt provides no quantitative benchmark table, dataset score, or measured baseline comparison.
  • The paper shows 3 ReAct loop variants where ACDL makes differences visible: base ReAct, no reasoning traces in action history, and query-based tool selection with tools placed later in the context.
  • ACDL models 4 common LLM API message roles: system, user, assistant, and tool.
  • The language supports nested time steps such as @T.I, which lets one specification describe outer chat turns and inner ReAct tool-use steps.
  • The claimed concrete output is a usable specification language plus 4 pieces of tooling: parser, renderer, VS Code plugin, and agentic skill.

Link