pith. sign in

arxiv: 2605.01920 · v1 · submitted 2026-05-03 · 💻 cs.AI · cs.CL· cs.MA· cs.SE

A Language for Describing Agentic LLM Contexts

Noga Peleg Pelc , Gal A. Kaminka , Yoav Goldberg This is my paper

Pith reviewed 2026-05-09 17:14 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.MAcs.SE
keywords LLM agentscontext engineeringagentic systemsdescription languageprompt structurecontext dynamicscontext visualization
0
0 comments X

The pith

ACDL provides a standard language to precisely specify and visualize the structure and dynamics of LLM contexts in agent systems.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

LLM agents rely on sequences of calls where each input context combines instructions, observations, and history, and the specific design of these contexts strongly influences system performance. No standard way exists to describe how such contexts are built or how they change across steps, leading to reliance on informal prose, diagrams, or code that fails to capture evolution or differences clearly. The paper introduces the Agentic Context Description Language (ACDL) with dedicated constructs for role sequences, dynamic content, time-indexed references, and conditional or iterative structures. These allow full, implementation-independent specification of context architectures, supported by both hand-drawn and rendered diagram forms. Adoption would enable consistent documentation, comparison of strategies, and clearer communication in both everyday work and research papers.

Core claim

We introduce the Agentic Context Description Language (ACDL), a language for specifying the structure and dynamics of LLM input contexts in a precise, readable, and standard manner, along with visualizations. 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.

What carries the argument

The Agentic Context Description Language (ACDL) and its constructs for role message sequences, dynamic content, time-indexed references, and conditional or iterative structure, which together describe LLM context architectures and their evolution.

Load-bearing premise

The listed constructs are sufficient to capture the full architecture of any LLM context independently of implementation details.

What would settle it

An example context from an existing LLM agent system whose structure or temporal dynamics cannot be expressed using only role sequences, dynamic content, time-indexed references, and conditional or iterative structures.

Figures

Figures reproduced from arXiv: 2605.01920 by Gal A. Kaminka, Noga Peleg Pelc, Yoav Goldberg.

Figure 1
Figure 1. Figure 1: ACDL visualization of three simple ReAct loop variants. Middle: base implementation. Left: no reasoning traces in view at source ↗
Figure 2
Figure 2. Figure 2: ACDL is concerned with describing the queries view at source ↗
Figure 3
Figure 3. Figure 3: ACDL Description of a simple multiagent main view at source ↗
Figure 4
Figure 4. Figure 4: ACDL descriptions of a multi-turn multi-step ReAct view at source ↗
Figure 6
Figure 6. Figure 6: ACDL Description of OpenCode main loop. Tools, view at source ↗
Figure 5
Figure 5. Figure 5: ACDL Description of OpenClaw main loop, without view at source ↗
Figure 7
Figure 7. Figure 7: Gemini Plays Pokémon Blue context structure vi view at source ↗
Figure 8
Figure 8. Figure 8: Six variations of the Mint context explored in our experiments. Across all variations, the user instruction message view at source ↗
Figure 9
Figure 9. Figure 9: Figure 7 from the tech report of the DeepSeek-v4 view at source ↗
Figure 11
Figure 11. Figure 11: ACDL description of the DeepSeek agent without view at source ↗
Figure 12
Figure 12. Figure 12: Tool-using agent. System messages frame the task; view at source ↗
read the original abstract

Large language models are increasingly used within larger systems ("LLM agents"). These make a sequence of LLM calls, each call providing the LLM with a combination of instructions, observations, and interaction history. The design of the encoded information and its structure play a central role in the quality of the resulting system, leading to efforts spent on context engineering. It is therefore critical to communicate the composition of the LLM context in a system, and how it evolves over time. Yet, no standard exists for doing so: context construction is typically conveyed through 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. To remedy this, we introduce the Agentic Context Description Language (ACDL), a language for specifying the structure and dynamics of LLM input contexts in a precise, readable, and standard manner, along with visualizations. 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 context, both in day-to-day communication and in papers. Tooling, examples and documentation are available at www.acdlang.org.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper introduces the Agentic Context Description Language (ACDL) for specifying the structure and dynamics of LLM input contexts in agentic systems. It defines constructs including role message sequences, dynamic content, time-indexed references, and conditional/iterative structures, along with associated visualizations that can be hand-drawn or formally rendered. The language is positioned as implementation-independent, and the manuscript demonstrates its use by documenting several existing LLM agent systems and variants, with tooling and examples provided at www.acdlang.org.

Significance. If the language's constructs prove sufficient and gain adoption, ACDL could establish a much-needed standard for communicating context engineering decisions in LLM agents, improving precision over informal prose, ad-hoc diagrams, or raw code inspection. The explicit demonstrations on multiple real systems provide concrete evidence of applicability and support reproducibility in research and development.

minor comments (3)
  1. [Abstract] Abstract: The claim that ACDL captures the 'full architecture of a prompt independently of any particular implementation' is central but would be strengthened by an explicit discussion of scope limitations or edge cases not covered by the listed constructs (role sequences, dynamic content, time-indexed references, conditional/iterative structure).
  2. [Demonstration section] Demonstration section: The paper documents several existing systems and variants, but providing a table or summary listing the specific systems, the number of variants shown, and which constructs were exercised in each would make the coverage claim easier to evaluate.
  3. [Language definition] The manuscript mentions that ACDL diagrams 'can be hand drawn on a whiteboard, or written in formal language which can then be rendered,' but does not include a concrete syntax example or grammar fragment in the main text; adding one would improve accessibility for readers.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive and constructive review of our manuscript on the Agentic Context Description Language (ACDL). We appreciate the recognition of ACDL's potential to serve as a standard for precisely communicating context engineering in LLM agent systems, as well as the concrete evidence provided by our demonstrations on existing systems. The recommendation for minor revision is noted.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper proposes a new descriptive notation (ACDL) for LLM agent contexts rather than deriving any quantitative result, prediction, or theorem from prior equations or fitted parameters. Its core claim is that the listed constructs (role sequences, dynamic content, time-indexed references, conditionals) suffice to capture context structure independently of implementation; this is supported by direct application to several existing systems, which constitutes external demonstration rather than reduction to self-citation or self-definition. No equations, uniqueness theorems, ansatzes, or renamings of known results appear in the provided material, and the work is self-contained as a language specification.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The contribution rests on the domain assumption that LLM agent contexts can be decomposed into the listed structural elements and that a dedicated language can describe them independently of code. No free parameters or fitted values are involved. The main invented entity is the language itself.

axioms (1)
  • domain assumption LLM contexts in agentic systems consist of role message sequences, dynamic content, time-indexed references, and conditional or iterative structure
    This decomposition is taken as the basis for the language design and is stated in the abstract as what ACDL captures.
invented entities (1)
  • Agentic Context Description Language (ACDL) no independent evidence
    purpose: Provide a precise, implementation-independent way to specify and visualize LLM context structure and dynamics
    Newly defined language and associated diagram notation introduced in this paper.

pith-pipeline@v0.9.0 · 5578 in / 1381 out tokens · 38528 ms · 2026-05-09T17:14:58.139466+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

  1. [1]

    2024.Neural Circuit Diagrams: Robust Diagrams for the Commu- nication, Implementation, and Analysis of Deep Learning Architectures

    Vincent Abbott. 2024.Neural Circuit Diagrams: Robust Diagrams for the Commu- nication, Implementation, and Analysis of Deep Learning Architectures. Technical Report arXiv preprint arXiv:2402.05424. arXiv

    work page arXiv 2024

  2. [2]

    2024.PDL: A Declarative Prompt Programming Language

    Ibrahim Abdelaziz, Kinjal Basu, Mayank Agarwal, Sadhana Kumaravel, Matthew Stallone, Rameswar Panda, Yara Rizk, et al . 2024.PDL: A Declarative Prompt Programming Language. Technical Report arXiv preprint arXiv:2410.19135. arXiv. https://arxiv.org/abs/2410.19135

    work page arXiv 2024

  3. [3]

    DeepSeek-AI. 2026. DeepSeek-V4: Towards Highly Efficient Million-Token Con- text Intelligence. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/ main/DeepSeek_V4.pdf. Technical report

    work page 2026

  4. [4]

    2025.Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini Team, Google. 2025.Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. Technical Report. Google. https://storage.googleapis.com/deepmind-media/ gemini/gemini_v2_5_report.pdf

    work page 2025

  5. [5]

    Guy Marshall, André Freitas, and Caroline Jay. 2025. An Evidence-Based Guidance Framework for Neural Network System Diagrams.PLOS ONE20, 3 (2025), e0318800. doi:10.1371/journal.pone.0318800

    work page doi:10.1371/journal.pone.0318800 2025

  6. [6]

    OpenCode: The Open Source AI Coding Agent

    OpenCode 2025. OpenCode: The Open Source AI Coding Agent. https://www. opencode.ai/. Open-source provider-agnostic coding agent for terminal, desktop, and IDE

    work page 2025

  7. [7]

    Peter Steinberger. 2025. OpenClaw: Your Own Personal AI Assistant. https: //www.openclaw.ai/. Open-source autonomous AI agent framework. Originally released as Clawdbot, November 2025

    work page 2025

  8. [8]

    Knut Sveidqvist and Contributors to Mermaid. 2014. Mermaid: Generate diagrams from markdown-like text. https://mermaid.ai/

    work page 2014

  9. [9]

    Xingyao Wang, Zihan Wang, Jiateng Liu, Yangyi Chen, Lifan Yuan, Hao Peng, and Heng Ji. 2024. MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback. InProceedings of the Twelfth International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2309.10691

    work page arXiv 2024

  10. [10]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Representations (ICLR)

    work page 2023

  11. [11]

    Naren Yellavula. 2024. PromptML: Prompt Markup Language. https://www. promptml.org GitHub repository

    work page 2024

  12. [12]

    Joel Zhang. 2025. The Making of Gemini Plays Pokémon. https://blog.jcz.dev/the- making-of-gemini-plays-pokemon Blog post

    work page 2025

  13. [13]

    2025.Prompt Orchestration Markup Language

    Yuge Zhang, Nan Chen, Jiahang Xu, and Yuqing Yang. 2025.Prompt Orchestration Markup Language. Technical Report arXiv preprint arXiv:2508.13948. arXiv. https://arxiv.org/abs/2508.13948 A Language for Describing Agentic LLM Contexts CAIS ’26, May 26–29, 2026, San Jose, CA, USA Appendices A The MINT Experiment We evaluated 7 context configurations that share...

    work page arXiv 2025

  14. [14]

    search" { U: env.search_results[@t] } Case

    models. The following is taken verbatim from the DeepSeek-V4 techni- cal report [3], which describes some of the context management strategies for their agent. DeepSeek-V3.2 introduced a context management strategy that retains reasoning traces across tool- result rounds but discards them upon the arrival of new user messages. While effective, this still ...

    work page 2026