Margarita is an agentic programming language and prompt template framework based on Markdown extensions, open-sourced by its developer, Banyango, in January 2026. This tool was developed to fundamentally address the "spaghetti prompt" problem, where prompt engineering with large language models (LLMs) becomes increasingly complex, resulting in massive and unmanageable text. Similar to how React divides and assembles complex HTML document structures into smaller components for web front-end development, Margarita introduces an architecture that allows users to modularize and assemble an agent's prompt fragments into file-based units. Users can maintain the standard Markdown format while embedding programming syntax, such as dynamic variable insertion, conditional statements, and loops, to dynamically construct instructions for the LLM. Existing prompt management methods fail to overcome the critical limitations of non-deterministic behavior and instruction omissions, which arise from relying entirely on the LLM's probabilistic response mechanism. In particular, in multi-stage agent workflows that follow complex logical flows, errors frequently occur where the model arbitrarily omits intermediate instructions or deviates from the defined output schema. Margarita ensures the rigor of execution by assigning direct script execution rules to the prompt. It supports state management and memory preservation mechanisms, explicitly controlling the information that the LLM must maintain and the tool calls it must perform at each step. This approach reduces the LLM's tendency for lightweight, impromptu responses and dramatically increases the predictability and consistency of system behavior. In domains that require high levels of regulation and precision, such as life science research or clinical trial analysis, Margarita can serve as a valuable companion framework. For example, when extracting specific protein domain information and drug interaction indicators from thousands of pages of genomic analysis papers or clinical trial result reports, Margarita's conditional expressions can be used to activate a precise extraction template only when specific biomarkers are detected, and a multi-sequence data set can be sequentially processed using loops in an agentic workflow. Furthermore, it inherently supports integration with locally running Ollama platforms or secure GitHub Copilot CLI infrastructure, enabling the construction of deterministic pipelines that can safely process sensitive patient medical information or confidential preclinical data on local hardware without the risk of external exposure. In addition, this framework provides a developer-friendly development experience. Unlike conventional prompt file management for plain text, the injection and rendering of variables can be validated through simple CLI tool commands, and debugging of agent scripts (.mgx) is also easy. This allows research data analysts and software engineers to easily perform version control and unit testing on the AI agent prompts they have built, increasing reliability throughout the project lifecycle.

Margarita is an agentic programming language and prompt template framework based on Markdown extensions, developed by Banyango and released as open source in January 2026.