June 26, 2026 · 2 min read · Jake Kirsch
Code Is the Ledger: Why Coding Agents Are the Right Way to Use AI at Work
A friend of mine works in finance. He's excited about AI, but wasn't sure where to start.
His team had a monthly report that took hours to put together, with contributions from many people across the organization. They ran a test: fed a bunch of Excel files and context into Copilot and recreated the report in minutes. The numbers matched. Everything looked right.
But then the real question surfaced: What happens when someone up the org chart asks how the number in cell B7 was calculated? Is "because AI said so" an acceptable answer?
No.
Here's a conclusion I've reached lately: the reason coding agents are taking off is two-fold.
1. They're really good now.
Not perfect, but they produce usable code in seconds to minutes, and they drastically accelerate any process that involves writing code. Why? There's an enormous amount of publicly available code to train on, and validating good code doesn't require a human in the loop. The feedback loop is fast and objective. This is what LLMs are best at.
2. Code is a ledger.
Think back to that financial report. Instead of feeding those Excel files into a chatbot, what if you handed them to a coding agent and said: "Write a script that takes these files and produces a report meeting these criteria."
Now the logic is written down. You can read exactly what's happening at each step. You can change it. You can look back at the full history of changes. Everything is auditable and verifiable.
Think about any process you do at work. If it needs to be verifiable: reports, reconciliations, data transforms, compliance checks. Have an agent write it as code. Run it, validate it, change it, repeat.
That's it. Simple as that. I'm doing this multiple times a day, both internally at Infraless Data and for clients.
Bonus: your private data never has to leave your environment.
A common concern with AI tools is sending sensitive or proprietary data to a model provider. The code-first approach sidesteps this entirely. Use fake or anonymized data when prompting the agent to write the script, then run that script on your real data locally. The model only ever sees the dummy data - your actual data stays in your environment and never touches an external API.
Jake Kirsch
Infraless Data