Travis Colbert

A key aspect to fulfilled programming in the age of agentic coding is retaining responsibility for the code that we produce. By we I mean us and the magic slo[tp] machines we employ. The moment we stop caring about the code is the moment we stop knowing the code. And not knowing the code leads to ambivalence.

There is an inverse reality, though: the less we know about the code, the less we can responsibly and confidently assert that our code can be relied upon by others. The less we care the less anyone else should care. No one wants to work on software that no one cares about.

But the paradox is this: even if we really want to care and take responsibility for our code, an LLM can produce such great quantities of functional-looking code so quickly that it is becoming less possible for the human guide to keep up. The nozzle velocity is too high. I've experienced PRs that were too massive to realistically read and understand in a single pass.

I'm a fairly modest guy. I know I've got limitations and I'm ok with that. One researcher says that people can remember seven plus or minus two chunks of information. I think I'm probably on the low end of that range. But, anyway, I'm fine with that. The point is: if my LLM is able to pump out volumes of slop, I have less of a chance of being able to confidently sort out the non-functional from the functional.

So, I recently set about making a simple tool that serves as a neutral third party - a disinterested advisor - in my session. It provides an unbiased, play-by-play of what your agents are doing to your code. I call it Enlighten, and I thought I'd share it with you.

Enlighten sits off to the side observing your code base. It is "unbiased" in the sense that it is fundamentally unaware of intent. However, as your coding LLM persists changes to your code or if a commit is made, Enlighten reads the change, analyzes it quickly, streams a concise, plain-English description of what just happened along with a guess of what the intent of the change is and what it believes is the likely next step in this session's development. If this unbiased third party LLM's summary roughly matches your intent, then perhaps you have greater reason to trust the changes.

The trouble I often experience is: at first pass every edit by your LLM carries the same cognitive weight as the next edit. So, it tends to be hard to separate the signal from the noise, the value from the slop. Enlighten's goal is to distill the signal and the noise so it is easier to pick out where your agents have inserted useless code or diverged from the original intent.

I thought this was a cool idea because this tool is essentially trying to figure out what your agent is doing just like the poor meatbag sitting in front of the screen is.