A few years ago, while looking into Typst, I came across a line: “The world has suffered LaTeX long enough.” Honestly, yes. I started trying to learn LaTeX in my first year at university. More than ten attempts, and every one ended with giving up. I got almost no immediate feedback from it. I could not get the hang of it, and I did not want to learn it anymore.
Then LLMs took off. I started with ChatGPT, manually copying and pasting over and over, stumbling my way toward a solution. When MCP appeared, I put together a LaTeX-MCP. Later, Claude Code could use Bash to compile and fix errors directly, and I stopped writing LaTeX myself. After all those years of trying, I finally got to use LaTeX. It was my agent that learned it.
Looking at it now, Markdown is enough for a first draft. Once it is written, I tell the agent which journal I am targeting and have it find a template and do the typesetting. But Markdown being enough only applies to drafting. Submission brings specific requirements for sections, equations, figures, tables, and references, and the journal’s template still needs to be used where required. AI summed it up for me: “Finishing the text of a paper means it is time to prepare the typesetting and submission. The journey has only just begun.”
In the age of agents, using—or learning—LaTeX may no longer mean learning its syntax. Our limited time is better spent understanding what it suits, which established solutions are available, and how to describe our formatting requirements clearly. As for the commands and how to fix compilation errors, leave those to the agent.
What is LaTeX?
LaTeX is a typesetting tool used for papers, books, research reports, and even slides for a group meeting. It organizes text, equations, figures, tables, and references in source files, then uses a template to generate a PDF through a process called compilation. Templates from universities and journals, together with established tools for mathematics, references, and figures, make it something of a common currency in research writing: an existing project can be edited further, and its source can be adapted to a different journal’s template rather than starting with a blank document every time.
LaTeX does require technical work, since adjustments that take a few clicks in Word may involve commands, configuration changes, and another compilation, while Chinese fonts, bibliography processing, and special template formatting can require troubleshooting first. But the ceiling is high: margins, fonts, section structures, equation numbering, figure and table notes, and bibliography styles can all be adjusted, with packages—components that add functionality—and custom commands extending what existing templates provide. This allows it to handle both short articles and books hundreds of pages long, while expressing complex formatting as reusable rules, although making full use of those capabilities traditionally meant investing substantial time in learning the tool.
Agents offer a different way to allocate that effort: researchers still decide the formatting requirements, while command lookup, source editing, compilation, and troubleshooting can be delegated without first becoming proficient LaTeX operators themselves. Time can instead go toward deciding what the manuscript needs, whether an existing template meets those requirements, and whether the generated result is right. The technical demands remain, but agents can now perform a substantial share of the operations involved.
What does it offer a social science paper?
For empirical researchers, writing and analysis often alternate: drafting the results section can send you back to add a specification, change a variable definition, or reconsider the sample after discussion. In a paper on a policy’s effects, for example, the main results, robustness checks, heterogeneity analyses, and corresponding figures need not be settled in the first draft. Each analytical revision requires updating the manuscript too. Copying output into manually formatted tables means transferring it again whenever the model changes; with LaTeX, Stata, R, or Python can export results to designated table and figure files that the manuscript includes. Updating through the same analysis program and recompiling can remove another round of copying and formatting.
Presentation also carries substantive information. A regression table in an empirical social science paper must communicate what each column estimates, what is controlled for, how many observations are used, and what the parenthesized values and notes mean. Which details repeat across columns, which belong in a shared note, and which column the text discusses are decisions tied to the argument. LaTeX can express those presentation requirements in table-generation and formatting rules, reducing manual rearrangement when a column or set of results changes. But a revised model still requires checking the headings, notes, and interpretation together: successful compilation does not establish that they correspond.
Revision and submission can also change what belongs in the main text or appendix: a heterogeneity analysis moves into the paper, a supplementary figure moves out, or a referee requests another check. Keeping results files separate from prose makes these changes primarily about where the manuscript includes each result and how it explains it, rather than copying the numbers again. Agent assistance can be scoped accordingly—for example, updating result files with a specified program and moving a table to the appendix while preserving the existing prose for the author to revise. For this kind of paper, LaTeX is useful because it connects with the analysis programs already in use, reducing repeated operations between analysis, presentation, and writing while researchers decide which results belong and what conclusions they support.
Established options
TeX Live / MacTeX + VS Code
This is a local LaTeX setup: TeX Live supplies compilation programs and common components, while macOS users can choose MacTeX, which builds on TeX Live. VS Code with the LaTeX Workshop extension handles editing and compilation. You can work with the .tex source beside its generated PDF, compile changes, and see the updated result alongside the text. Manuscript files, figures, tables, and references stay in your project directory; keeping the complete project lets you move computers or hand it to someone else for further editing.
This combination suits people who want to manage their files and work in a local editor, particularly when analysis already happens in Stata, R, or Python and exported tables and figures can go straight into the manuscript project without separate uploads. Once the environment and required components are available, much editing and compilation can happen offline, though fonts, packages, and template requirements need local attention. Keep the editor choice simple: if you already use VS Code, there is no need to investigate a new collection of tools just for LaTeX.
Overleaf
Overleaf is an option when you would rather avoid setting up local compilation, or your collaborators already write in the browser. Editing, compilation, and PDF preview live in one online project; a template gives you a starting point, and collaborators can edit and discuss the same source without repeatedly exchanging archives. That matters in joint writing, where working in one project removes the preliminary task of determining whether everyone has the same manuscript version.
Overleaf provides source and visual editing. The latter reduces some command entry, but complex templates or unusual formatting may still require source changes. Using the platform also means uploading the manuscript and supporting materials, with collaboration, history, and compilation resources depending on the current plan. If materials should not be uploaded, or a local workflow already works well, moving a project just to change tools adds work. Collaborators who mainly use Word’s tracked changes also need an agreed delivery format: conversion tools exist, but complex equations, tables, and citations still need checking afterwards.
Agent Driven
The core of Agent Driven is decoupling content production from typesetting. Researchers can focus on the question, argument, and expression, drafting in whichever tool they find convenient and offloading the work of turning that content into a suitably formatted LaTeX document to an agent. This is what it means for Markdown to be enough for a first draft: writing a paragraph need not also involve placing it in a template, and preparing a results table need not immediately interrupt the work with questions of width, font size, and page breaks. Establish the content first, then give the agent the manuscript, target journal, and formatting requirements together.
In this workflow, LaTeX provides the typesetting system while the agent handles fitting content into a template, adjusting figures and tables, and resolving compilation problems, reducing the need for researchers to keep switching between writing and formatting. Revised text can go back to the agent, and a different submission template can be applied around the same content. Researchers still review the final document for unintended content changes, correct figure and citation relationships, and compliance with formatting requirements, but more attention can stay with the paper itself while the agent handles expressing those requirements in LaTeX.
Give the agent the content and requirements
Start by defining what is being handed over. A substantially finished manuscript can be assigned for formatting while its text and research content remain unchanged; an unfinished paper can be handled one chapter, table, or group of figures at a time. Separating writing from typesetting still allows them to alternate: keep drafting while settled material is formatted, then update it as work progresses. Identify the manuscript version and scope so that the work stays anchored to the same content.
Explain the intended use next. A draft for collaborators, a working paper, and a formal submission can have different requirements. If a target journal is settled, supply its submission instructions, template, and other explicit requirements; otherwise agree on a readable, editable layout to adapt later. State the deliverables too: perhaps a PDF for review now and editable source for collaborators afterwards, or separate anonymous and title-page documents if the submission system requires them. These decisions affect how the material is organized.
When a template exists, ask the agent to arrange the content within its existing layout rules rather than imitate a screenshot from scratch. Establish its source and whether it is the journal’s current version. If a template has not been found, ask the agent to locate it and explain the basis for selecting it; if none is available, explicit formatting requirements can guide a provisional layout without inventing an “official journal template.” Requirements such as abstract length, anonymity, or figure placement may appear only in submission instructions, so a template compiling successfully does not establish compliance with everything else.
Explain how the manuscript and supporting material correspond, particularly when the project contains several versions of the results. Identify which figure belongs in each section, which table is current, and whether titles and notes are settled; that information is more useful than teaching the agent an image-insertion command. A newly exported regression table can replace the old one using the existing layout while preserving its values, column labels, and notes. This is a formatting update, not a request to re-estimate the model. Missing material can be identified and supplied later without turning typesetting into the invention of research results.
Separate reference identity from reference appearance. If a bibliography database or prepared reference list exists, ask the agent to reuse those records, match in-text citations to entries, and apply the required style. If the draft contains only author–year notes, also supply titles or other identifying information, especially where an author has several works in the same year. Uncertain citations can be collected for clarification rather than completing a convincing-looking bibliography by guessing authors, years, or DOIs. Formatting and source identification are distinct tasks.
Distinguish fixed requirements from adjustable layout choices. Institutional fonts, margins, and line spacing may be prescribed, while column widths, figure sizes, and pagination can be adjusted without changing content. For wide or long tables, explaining how readers should use them is more actionable than asking for something “more elegant”: columns may need to remain together for comparison, or a table continuing onto another page may need repeated headings. Decisions about which results belong in the main text or appendix concern the paper’s structure and can be made by the researcher before the agent implements them.
Once the first version is ready, open it and request changes based on the visible result rather than anticipating every possible issue upfront. Identify a page, table, or heading, explain what is hard to read, and describe the desired change—for example, “the notes under the second results table are cramped; keep the font size and see whether column widths and spacing can be adjusted.” Working on one class of issue at a time makes the differences easier to assess. While the text is still evolving, specify whether a round updates content while retaining layout or changes formatting while preserving wording, so that previously settled material is not needlessly altered.
The final handoff can include material that remains useful afterwards: the generated PDF, its complete editable source, and a clear account of outstanding issues. The PDF shows the current result; the source supports further revisions, collaboration, or adaptation to another submission target. Keep the relationship between that output and the underlying content clear rather than retaining only a finished-looking document. Researchers ultimately check completeness, matching figures and references, and compliance with requirements, while the agent handles LaTeX organization and commands. The handoff remains about content, requirements, and deliverables.
What Marx Max can help you do
Marx Max connects existing manuscript materials to agent-assisted typesetting. Researchers describe their requirements in natural language, and the agent can help prepare editable LaTeX source and generate a PDF in an available compilation environment. A Markdown draft, exported figures and tables, and reference materials provide a starting point: specify the template and scope, then work toward a document that can be read and revised further. Researchers who do not want to write LaTeX can use its typesetting capabilities while retaining a familiar drafting workflow.
With a university or journal template already available, the agent can work within it to arrange sections, equations, figures, tables, and existing references, then generate a document for review. Further requests can identify a table that is too wide, an unsuitable heading level, or material needing a different arrangement; the agent can edit the source and compile again. Researchers describe how the paper should appear while the agent handles expressing those requirements in LaTeX. Submission instructions beyond the template can be supplied as part of the task too.
Marx Max’s LaTeX capabilities can inspect existing compilation tools and test basic usability with a small document, reducing the work of determining whether the environment is ready. Existing tools are preferred, and missing components can be addressed based on the findings. When an environment works but a manuscript fails, the agent can read logs to identify missing files, font mismatches, or source problems and work on the specific cause. Troubleshooting can stay focused on the paper without requiring the researcher to become proficient in environment configuration first.
For researchers already doing analysis in Marx Max, another practical use is bringing results into the manuscript. Text, analysis programs, and figures can belong to the same research project, with the agent using specified output files and the paper’s existing title, note, and layout conventions. A completed supplementary analysis, for example, can be assigned for insertion into the appendix while preserving its values and notes and updating the relevant in-text reference. Further analysis can be requested separately. Connecting existing output to writing reduces repeated copying between conversations, result files, and the paper.
Revisions can continue around the same manuscript: identify updated text, specify replacement result files, or provide a new template and requirements when the submission target changes. Each round can have a limited scope, such as replacing results without rewriting prose or changing layout while preserving settled content, followed by a newly generated PDF for review. For papers that repeatedly return to analysis and revision, typesetting can follow research progress without requiring the entire manuscript structure to be explained again each time.
The useful deliverable includes both readable output and material for further work. Ask the agent to retain the generated PDF, corresponding LaTeX source, supporting figures and references, and an account of outstanding requirements. The PDF supports discussion and review; the complete source can go to collaborators or be edited in other LaTeX tools, keeping future workflows open. Changes remain inspectable in actual files rather than existing only as explanations in a conversation.
Across the research process, Marx Max provides a way to connect materials, analysis, writing, and typesetting: researchers can continue discussing content and working with existing results in a project while delegating document operations to the agent. The research question, argument, and interpretation still deserve attention, while arranging settled content in a template, generating documents, and implementing formatting feedback can take less direct effort. Separating content production from typesetting then leads to a manuscript that can be located, checked, and edited further.