Coding agents are becoming the most capable tools for repeatable knowledge work. You can use files, folders, and instructions to turn them into systems that automate hours of office work.
Until recently, organizations had two obvious ways to use AI to automate knowledge work. They could give employees a general-purpose chatbot, like ChatGPT, and leave it up to their staff to figure out how to make it work, or they could build or buy software designed around a specific process and make it work for their workflows.
Both approaches can work, but they leave a large gap.
Chatbots are flexible, but the workflow often remains in the employee's mind and might vary from person to person. Each person has to reconstruct the process through prompts, and the quality of the result depends heavily on how well they use the tool. Purpose-built software might be more consistent, but it can be expensive to buy, slow to develop, and difficult to change as the underlying work evolves.
Coding agents like Codex and Claude Cowork offer a useful third option. Unlike a chatbot, they can carry out a long, multi-step workflow: reading source files, following detailed instructions, creating and editing deliverables, checking their own work, and taking follow-up actions based on whatever they find. Unlike purpose-built software, they do not require an engineer to translate the entire process into a software application before it can be used.
For many file-based knowledge-work tasks, this gives subject-matter experts a faster and more flexible way to automate repeatable work while still producing consistent, reviewable outputs.
In this article, I'll show what this looks like through a simple example: turning two trial-balance files into a 90% done Excel variance analysis workpaper in a few minutes.
The coding agent combines the two trial balance files as tabs in an Excel file, creates a comparison tab that uses Excel formulas to link to the source tabs, applies materiality thresholds to variances, and generates follow-up questions for significant variances.
A better automation option: files and instructions, not software
The key shift is that you don't have to turn every repeatable process into a traditional software product before it can be automated. For many file-based workflows, the process itself can live in a set of plain-text instructions and examples that a coding agent reads and executes against the source material.
Here's how the setup looks:
- The workflow = instruction files, thresholds, examples, validation rules
- The executor = Codex, Claude Cowork, etc.
- The inputs = spreadsheets, PDFs, CSVs, documents
- The outputs = normal business artifacts like Excel files, memos, reports
Automation example - Variance Analysis
In this article, I'm going to walk through using this file-system based approach to automate a task. Starting with a very simple workflow, here are the exact steps I went through to automate the process to turn trial balance CSVs to a real Excel analysis with the data compiled, linked, analyzed, with questions ready to go for further research.
The basic workflow looks like this:
- Compare 2 years of year-end account balances in Excel.
- For accounts with large or unexpected variances, review additional data or work with the client to determine what contributed to the change.
- Determine if the change is reasonable or if additional explanation or testing is needed in any area.
For this example, I saved the task instructions as a file, it describes where to find the source files, what workflows to perform on them, and tells the agent what thresholds to use for generating questions.
Request: 
How to think about building your own workflows
For most of the workflows I've built, this comes down to two main files:
- AGENTS.md - A project-level file describing the overall project goals.
- SKILL.md - To give details about the step-by-step workflows the agent will execute.
Other options include having a folder of examples showing the agent what good output looks like, a schema or template to follow, and any reference material the agent needs to do the job.
AGENTS.md
The AGENTS.md file describes the kind of work performed in the workspace, I keep this file as short as possible but at least they will lay out where skills files or reference files live, and where files should be saved to.
SKILL.md
There is just one SKILL.md file for this basic task, called variance-explainer. It's the file that actually describes step-by-step workflow or repeatable task I need to automate.
While I described the task above as "compare 2 years of year-end account balances in Excel," the reality is there are actually a lot more tedious steps to the task than what is implied.
This file instructs the agent to perform all the steps a person would go through:
- detect columns and normalize values
- build a key of values to compare (i.e., account number)
- create comparison fields with Excel formulas
- apply thresholds to generate analysis notes and client questions
- apply sorting to the resulting analysis
- create a summary tab to document the process and analysis results
- filenames to follow for generated output
To ensure the output is consistent and repeatable, the SKILL.md file also specifies what Excel output should look like:
Summarytab that includes details about the:- source files for the analysis
- purpose of the analysis
- thresholds used for generating research questions
- details about the data in each file
- data validation details
Prior_Year_TBtab that includes the exact source file used.Current_Year_TBtab that includes the exact source file used.Comparisontab that includes the normalized account comparisons with formulas linking the account numbers to the source tabs with SUMIF formulas. It also tells the agent to apply a dollar and a percent threshold to the account variances to determine when to create follow-up questions to answer for the large variances.
Walkthrough of Codex completing the task
The short summary is that the agent was able to produce a high-quality Excel spreadsheet that has a summary of results tab, source tabs for each input file, and a tab with accurate formulas that reference the source files to perform the annual comparison. It also applied the threshold accurately and included appropriate follow-up questions that should be answered before the workpaper is complete.
Honestly, I haven't found any issues with this Excel file so far and this is very close to the spreadsheet I would create to prepare questions to answer with additional research or questions to management.
Here are a few screenshots of the end result:
Summary tab: 
Comparison tab: 
Example Formulas from Comparison tab: 
Lessons
The output is still a first draft but it is a very high-quality first draft. The spreadsheet it produced is close to what I'd build by hand, it included accurate formulas, source tabs, thresholds applied, questions ready to research. But those questions are a starting point. The auditor still decides which variances are reasonable and which need follow-up. The agent helps with the mechanical work of getting to that decision quickly and setting up the workpaper.
Iterating the skill file took a lot longer than I anticipated because everything I could do without thinking needed to be described, like detecting which column holds the account number, normalizing values that don't match between files, deciding what size of variance is worth a question. None of that is hard for the models available through Codex and Claude Cowork, it's just tedious to spell out in detail. That's the cost of this approach. However, once the steps are on paper, you own them and modify as needed.
If you've got a process you can describe step by step, and a lot of knowledge work is a lot of detailed steps, you can probably put the process in a file and hand the execution to whichever model or agent is best that month.
Here's a link to the files and project setup I used: https://github.com/scottlabbe/codex-business-workflow