Development often starts with a defined outcome to build; research uses successive attempts to determine what to do next. For a login feature, the intended user actions and behaviour can be specified before asking AI to implement and verify them. Investigating an economic phenomenon rarely allows every analytical step to be settled in advance. Data may differ from expectations, a model may raise another question, and an explanation may need reconsideration. Each calculation matters both for the step it completes and for the evidence it provides for the next attempt.
Before building Marx Max, I had been using Claude Code and Codex, and I value their strong coding capabilities. Those capabilities can already help researchers write programs and run calculations. The further question is how AI should collaborate as the direction of analysis changes. A researcher may begin with a plot, compare specifications, then revise variables or add checks in response to the results. Assigning an entire project a fixed route at the outset can bypass moments that need human judgement. Collaboration in research needs room for researchers to revise the plan during exploration and for AI to continue under the updated requirements.
Marx Max brings discussion, editable code and calculation results together to support that exploration. On the left, researchers discuss the current question and next attempt; on the right, they inspect actual output, revise code and rerun calculations. Earlier work can remain available for comparing results under different conditions before choosing a direction. Sections therefore give observation, discussion and analytical revision a practical place in the process. AI assists with programming and calculation while researchers decide what comes next from the evidence: that is the research-specific rationale for this arrangement.
Research requires explicit conditions beyond execution
In software development, code commonly forms part of an application, service or tool. In empirical research, programs process data and calculate evidence, while the research contribution also includes an interpretation supported by that evidence. Development likewise requires reasoning and verification. Research nevertheless has particular conditions that need explicit treatment: where data comes from, who or what is studied, how variables are constructed and how a calculation relates to the question. Successful execution cannot establish those conditions by itself.
For example, the same regression program may support different comparisons when the sample, variable definition or set of controls changes. Researchers need a rationale for each adjustment and must assess whether differences in output arise from specifications, samples or other processing conditions. The work is not merely making a program executable: it uses calculation to examine a research judgement. Without connecting final code and output to these conditions, the researcher still has to organise the relationships before continuing.
Research-oriented optimisation therefore extends beyond generating code to explaining research conditions, inspecting calculation and comparing results. Marx Max brings requirements into the calculation workspace and retains inspectable, editable code and output so these activities can proceed together. Efficiency should be assessed accordingly: after AI assists with programming and calculation, can the researcher understand the result, revise the analysis and continue more easily, with less effort across the complete task?
The concrete differences become clearer when a defined development task is compared with exploratory empirical research:
| Aspect | Development with a defined outcome | Exploratory empirical research |
|---|---|---|
| Starting point | Specify how a feature should work, such as login and permission checks. | Pose a question and examine possible explanations through data and analysis. |
| Basis for the next step | Organise implementation, testing and revision around confirmed requirements and acceptance conditions. | Use observations and model results to choose the next comparison, revision or additional check. |
| AI’s work | Write code, investigate problems and run tests for the given outcome. | Perform the currently specified processing and calculations, show results and continue under the researcher’s revised requirements. |
| Human participation | Confirm requirements, make design choices and assess expected behaviour. | Assess samples, variables and methods, then choose the next step from the evidence. |
| Evaluating the result | Determine whether behaviour and quality meet the established requirements. | Determine whether specifications and evidence support an interpretation, with clear conditions and scope. |
| Material to retain | Implementation code, test records and requirement changes. | Data sources, sample treatment, code, outputs under different specifications and reasons for analytical choices. |
Turning a request into a process you can inspect
Consider comparing a baseline model with a model containing additional controls. The question concerns how the core coefficient and its uncertainty change under clearly specified conditions. A meaningful comparison requires identifying the dataset, included observations, variable definitions and standard-error method. Those conditions determine the calculation; a tool should not fill them in merely to produce a result sooner.
In Marx Max, specify the baseline on the left and request its calculation code and actual output on the right. After inspection, add the selected controls while retaining the original model. Check more than the final coefficients: the input file, sample selection, variable list, intercept and standard-error method also matter. Where a condition needs revision, identify the relevant step and continue from the correction rather than repeatedly describe the entire research background.
A specific issue arises when observations drop out because the additional controls contain missing values. Differences between the models then involve both the variable set and the sample, so the coefficient change cannot all be attributed to the controls. Inspect effective observations and perform the required comparison on a consistent sample where appropriate. This does not replace judgement about the controls or interpretation, but helps remove an unnecessary ambiguity from the comparison.
Keeping requirements, code and output together helps the researcher inspect these relationships. AI implements confirmed conditions, performs the requested calculations and helps organise the comparison. The researcher decides whether the conditions are appropriate, how to interpret changes and whether further checks are needed. With an explicit basis for both contributions, AI output becomes more usable in the next stage of analysis.
Staged execution supports comparison and revision
Research analysis includes loading data, constructing variables, selecting samples, estimating models and organising results. These stages establish conditions on which subsequent interpretation depends. Results may also prompt reconsideration of a variable or specification, so analysis does not always follow a settled script in one pass. Presenting stage-specific code and output helps researchers inspect conditions while the relevant material remains clear and decide whether to continue, revise or extend the analysis.
For data preparation, inspect names, units and observation coverage before handling missing values and selecting the sample, then confirm the treatment before estimation. Corrections can happen before many model results have been produced. Each stage needs pertinent checking information, such as observation counts, variables used and actual errors. A statement that processing is complete is insufficient to establish that the requirements were followed.
Marx Max’s sections of code and results support this arrangement. Confirmed parts can remain, relevant parts can be revised, and dependent calculations can then be updated. Staging need not require approval of every line. Stable repetitive calculations can be grouped; choices that change the sample or interpretation deserve confirmation. Concentrating checks on consequential decisions can reduce both rework and unnecessary communication.
Once an analysis is stable, establish a clear execution order and inspect the complete process from a defined starting point. Sections support exploration and revision; a complete run checks dependencies. These arrangements serve different stages, giving researchers flexibility during exploration and a consistent calculation process for formal results.
Collaboration requires direct participation in the work
Researchers need to participate in changes to code, materials and results, beyond approving or rejecting AI’s final output. Some changes are easier to describe in natural language; others are more accurately made directly. A researcher who knows how a variable should be constructed should be able to edit and rerun the relevant code without turning that adjustment into repeated explanation and guessing.
On Marx Max’s right, researchers can inspect output, edit code and rerun calculations; the conversation on the left establishes requirements and discusses subsequent work. When continuing a task, explicitly identify changed conditions and ask AI to work from them. Human adjustments enter the calculation, and AI’s contribution remains usable by the researcher. An edit should not become a reason to regenerate the entire analysis unnecessarily.
Responsibilities should reflect research tasks. The researcher determines the question, variable meanings, rationale for comparisons and conditions for interpretation. AI can assist with repetitive programming, specified calculations and organising output; proposed specification changes bring their reasoning back for discussion. The arrangement permits direct human participation alongside substantial AI work. Research-oriented collaboration should use coding tools’ strengths while keeping calculation directed at the author’s question.
Repeatable calculation gives output continuing value
Research results often need to be presented, revised or reviewed. Researchers must know which input, program and settings produced an output to understand later changes. AI explanations can assist understanding, but retained code, materials and actual results provide the basis for further inspection and repeated use.
Marx Max retains code and results in the calculation workspace for saving and revision. Saving files differs from retaining memory state: kernel variables do not automatically return after reopening the application. For formal results, run prerequisite steps in order and confirm the intended data and current code; make settings explicit where randomness is involved. Reproducing output establishes repeatability under those conditions, while interpretation still depends on method and evidence.
This distinction makes the collaboration more useful. Researchers can identify which outputs are ready to reuse and which prerequisites must be prepared. Retaining programs and results reduces repeated writing and searching; checking execution conditions establishes their relevance to the next task. Defined contributions allow generated work to serve an ongoing project.
Suitability for research must be tested through tasks
Begin evaluating Marx Max with a defined analysis: identifiable material, stated model conditions and clear deliverables. Observe whether AI follows the requirements, whether code and output are readily inspectable, whether revisions permit continuation through the correct steps and how much effort the full task saves. The comparison concerns the complete working process, rather than treating generation speed as overall efficiency.
Completion criteria for a software-development task cannot substitute directly for evaluating a research project. Beyond execution, a research workspace should make samples and specifications clear, calculations comparable, human edits usable in subsequent analysis and programs repeatable under explicit conditions. These requirements are compatible with general coding tools. They explain why we organise our product around research activities and give discussion, calculation and revision corresponding places.
Combining conversation with calculation, editable code, inspectable results and checks of analytical stages and conditions are Marx Max’s concrete research-oriented design choices. They support stating research requirements, human participation and examining evidence, with the shared aim of making powerful AI coding and calculation capabilities more convenient for research. Their value must be realised in actual use: researchers can continue comparing, interpreting and revising the work while devoting more attention to questions and methods.