About Agentic Coding

After spending so much time using agentic coding tools, it’s become clear to me that although AI is a multiplier, it is in the end capped by the human and our capacity to understand and direct its output.

I’ve gone pretty far in the spectrum of AI tools using orchestration and what not but I’m sure many people have gone farther, with their own intricate workflows/skills/mcp/roles/docs/tests/commands and a list of other things.

A lot of advocates in the AI coding space and people who actually deserve our respect like Steve Yegge, Thorsten Ball(author of 2 popular compiler books), the creators of OpenClaw, Pi Coding agent, Flask, Redis, say that AI is the biggest shift in software engineering. However, even in this camp there are 2 distinct lines of thinking, the first 3 say token intelligence will overcome all intelligence, while the latter say that it’s an amazing tool but still needs direction.

I’m close to the latter camp but still not as optimistic as them, even though I acknowledge we have no idea what can happen in a year.

Pre-AI there was always someone who actually wrote the code, understood it, cared about it. Now, a lot of people say you don’t need to see the code as long as it works, tests pass, and you can manually verify.

My personal experience with the no-hands approach was that as the codebase size increased I quickly lost motivation to work on the project as my understanding of the product and the ground-reality(code) became vastly detached. This experience has stayed the same for the last 6 months with moderate sized, moderate complexity projects.

In the end, humans are responsible and accountable for taking ownership. AI can’t do that. So it (still) pays to actively review, validate, and work on the code so you actually know what’s going on.

I’m still very much FOR using agentic tools but not to the capacity everyone else is raving about. I’m okay with using a dumber model, deliberately steer it multiple times, and review every piece of work it does. This works really well for me in terms of learning new things as well as producing meaningful output(without spending my usage limit on Opus). I also keep my motivation alive because I know what’s going on.

What’s annoying right now is seeing the push to fully orchestrate AI agents to do your work. I can probably bring up 50 job posts that mention that they want someone who can orchestrate agents and generate megaslop(my word not theirs) to push the boundaries of their slop factory. (ref: clickup layoffs)

2 things might happen: models will get so good that I’ll regret posting this or the industry will ground itself to the current state of models and get all the benefits AI has to offer without the side-effects of having a brittle mental model of the work you’re actually doing.

Whatever does happen, we have to carry on with life so stressing won’t actually help. Do whatever you can. For me, this has been a great time to learn deeply about fundamentals thanks to AI.

If you don’t have the money to use these tools, know that there are free tools with free tokens like OpenCode that you can use to get started.

And lastly, if you think tokens are heavily subsidised, please watch the Pragmatic Engineer podcast with Dax, OpenCode founder, he explicitly mentions that inference is hugely profitable even in the current climate where hardware is so scarce. So, it’s unlikely tokens will become expensive.

The real question is whether models will become much much better than today. Only God knows the answer to that.

Would love to know if you have a different view/workflow that actually helps you ship meaningful outcomes without the problems I mentioned here. Thanks for reading till the end :)