and this is a simple RAG context, where each query is fully independent from the previous ones (fresh context)
our language, ACDL (agent context description language) makes it easy to describe the structural semantic composition of an agentic prompt as it evolves over time. for example, here is ReAct context with tools at the end and history that includes reasoning traces:
Users appreciate ACDL because it helps describe agents, prompt structures, and system designs more clearly and effectively.
No Digg Deeper questions have been answered for this story yet.
we have been using ACDL internally in our group for some time now, and i found it to be really helpful when talking about agents and their implementation. and documenting the context in ACDL many times makes you understand it better.
Next time you’re publishing an agent and want to describe how it works, comparing your design to another system, getting teammates up to speed with shared vocabulary, or debugging what's really in your context window - consider trying ACDL!
Take OpenCode: system prompt + env info, a summary of compacted history, then each past turn replayed (user msg → tool calls/results → reply), then the current turn with its mode reminders (plan/build) and tool steps so far. ACDL describes it neatly.
Open tooling, all at https://acdlang.org: • Web editor + visualizer • VS Code extension • Claude Code skill (generate Python from ACDL spec) • Syntax reference, examples & tutorial
Joint work with @yoavgo and Gal Kaminka. You can also read our CAIS ‘26 paper: http://dl.acm.org/doi/10.1145/3786335.3813126 Describe your agent in ACDL, we’d love to see it!