pandoc

The agent constructs documents by writing Markdown files containing YAML metadata block headers, standard Markdown tables, and citation keys. It then compiles the document into a PDF in a single step using command-line arguments like --citeproc, --bibliography, and --pdf-engine=tectonic. This configuration works exceptionally well because Pandoc automatically coordinates complex formatting tasks such as bibliography pagination, section numbering, and table rendering without low-level TeX programming.

The command-line interface operates with high determinism. Successful compilation processes return an exit code of zero, which the agent uses to verify syntax integration. Because the official documentation provides clear examples matching standard markdown patterns, the agent successfully configures and executes the compiler on its first attempt, avoiding repetitive trial-and-error loops.

However, document verification works poorly. Because Pandoc outputs binary files like PDFs, the agent cannot programmatically inspect the visual rendering, page breaks, or text alignment. Deficiencies in the output layout require external multimodal tools or human review to be detected, leading to a weak verification workflow.

Additionally, multi-file asset coordination works inadequately due to fragmented tool dependencies. Since Pandoc cannot generate images or charts internally, the agent must write and execute separate Python scripts to generate visual figures prior to compilation. If an external figure script fails or produces a flawed file, the compiler still completes successfully with exit code zero but produces a visually incomplete document.

Pandoc is a highly suitable fit for agents that need to programmatically compile plain text data and structured headers into standard document formats like PDF, HTML, or Docx. It is ideal for standardized reports, academic papers, and briefs that rely on consistent structural layouts, automated bibliographies, and table-of-contents generation.

It is a poor fit for tasks that require precise visual positioning, manual typographic design, or real-time layout adjustments, as the agent has no built-in mechanism to preview coordinates or styling. It is also unsuitable for complex media pipelines that demand unified dependency tracking, as the agent must orchestrate visual asset generation entirely outside of the compiler.

The diagram shows how the agent coordinates asset generation and metadata compilation inside the document writing loop.