Your agents need write access
AI agents need write access to do actual work. Insights aren't enough.
There are huge swaths of the AI industry that have become consumed with using AI to deliver insights. Insights are fine, but if you actually want to change the way you work, to use AI to massively boost productivity, to make more money, and to have more time, using AI just for read-only summarization and insights is just the tip of the iceberg.
One of the big areas where AI has changed the industry is with coding. Now, if we limited agents just to summarizing code or to answering questions about code, would we see the massive changes we see today? No. The big reason that AI has transformed coding is because with version control, we can safely let agents write. If something goes wrong, if an agent deletes a bunch of files or undoes a bunch of work or introduces a bug, the full history of how it happened is right there in the version control system, and undoing it is easy. There’s no magical other place we need to look. We don’t need to go digging through logs. We know what changed over time, and we can very easily inspect that, or have an agent inspect that, and roll it back.
Now, if we look at lots of other areas outside of coding where people are trying to deploy AI, a lot of them are just living on this summarization and insight surface level.
Look at the financial space. How many different tools out there exist where you can plug in all your financial data and ask AI a bunch of questions? They all kind of suck. They don’t do much of value. I don’t need to ask daily questions about my finances. They don’t change that fast. What I would really love is if AI would do the books, send invoices and manage my vendors, and to do that, it needs to be able to write.
In a lot of other areas, because we don’t store data in a way that has the history natively built in, like how software version control does, we have become content with letting AI just read data and give summaries. That can be powerful, but that’s maybe giving you 10% of the ROI.
The thing holding back a lot of businesses from growing, from doing more with the people they have, from becoming more productive, is not that business’s inability to make decisions. If we limit AI to read-only workflows, we might be able to help businesses make slightly better decisions, but we cannot have those decisions carried out faster because, once the decision is made, people need to go and do the work. If AI is only summarizing and is only giving insights, it is not doing the work.
What we need to do is to enable AI to actually do work and to write, we need to bring the same kind of history-built-in safety we have with coding workflows, to the rest of work. That’s why I made Stellar Jay. Stellar Jay was inspired in many ways by Git, by the idea that everything has a version history.
If we instead of thinking of a business as a set of facts as they exist today, think of a business as a series of events that have happened over time, and we capture that, we can unlock the same kind of workflows we see in coding, in all sorts of other areas of business.
If you’d like to check out Stellar Jay yourself, you can here.
Originally published on X.