Travis Colbert

I don't even bother putting "I" in the sentence anymore. The agent had come up with a really nice UI workflow for a recent project. Everything looked and worked well except there was a nagging little Reacty refresh on the components in the new workflow that needed to be ironed out.

"No problem," I thought. "The agent will fix it."

I can't remember the exact prompt but I thought it was helpful and descriptive enough. The truth is that I didn't fully understand the scope of the problem at the time so I just described the bug and expected the agent to work through the problem. This is probably fodder for another article in the future. Anyhoo...

The Tight Loop

The agent dutifully started working on the problem. I could see its "thoughts" scrolling past. It checked all the foundational stuff, re-examined the components, and seemed to reason out the functional flow.

It reassuringly reported that it had a clear picture of the codebase.

"Great! This will be licked in no time", I thought to myself.

More deliberating and deep thinking.

I glance over to its train-of-thought.

More tokens burning... then: "Wait!..."

You've seen that emphatic "Wait". It's way too anthropomorphic, isn't it? It's supposed to evoke the idea that the agent seemingly realizes something new. Something it missed? Whatever it was, the agent needed to regroup and deliberate more. My eyebrow raised as my hopes grew that the agent was about to do something magical.

More tokens smoke on the altar... then another Aha! moment. I wish I would have kept this session - that's a little bit of a disadvantage of the CLI agent - but, again, the agent appeared to have discovered something important. Why is the agent uncovering course-changing details after the second reconsideration? Both of my eyebrows raise.

More reconsideration... more carbon... "I now see the issue", the agent triumphantly announces as it explains the new framing.

I let the agent continue to the point that I was convinced it had descended into a token-burning death spiral:

  clarity --> deliberation --> re-framing
     ^                             |
     |                             |
     +-----------------------------+

I had to eventually intercede and terminate the task. It was a Friday afternoon and I figured I'd give myself a chance to think through the problem myself (what a crazy thought) and fight another day on Monday.

Sure enough, as the weekend progressed some background process in my brain realized that this was probably nothing more than an event-bubbling issue that should be quite easy to correct. On Monday I prompted the agent with my theory and, in no more than a minute or two, my agent had corrected the problem and a PR was up.

The Moral of the Story

Don't worry humans. We're all going to be ok.

I've found that I've spent an unhealthy amount of time in low-key despair mode. This anecdote is helping me snap out of it.

Although our world as software engineers has completely imploded there's still room for human intuition. There will always be a need at some level or another for human guidance informed by experience.