Thread by @nateberkopec

I'm sold. Agentic coding is the future of web application development. There is no going back. Close the editor. Open Claude.

Your job is now to manage, review, corral and improve a relentless junior dev who is working on 6+ PRs in parallel.

If your mental model of LLM-mechanized coding is just tab-autocomplete in Cursor, you should try closing your editor for a week and only work out of Claude Code w/a parallel workflow using git worktrees (many tools now do this). Work on improving the agent if output is bad.

I'm less sold on agentic coding in the following contexts:

1. Libraries
2. Extremely large open-source codebases with long history (i.e. the METR study)
3. Concurrent/distributed systems
4. Probably more I can't think of. It's magical autocomplete, not a panacea.

I can see really high-skill, high-taste shops like 37signals struggling to adopt this because their very good taste means they are repelled by the slop. Bigger companies and small startups/indies will have an easier time at it, out of necessity.

In an agentic coding workflow, the LLM gets your 80% of the way there, and then has to work side-by-side with you to turn that "80%, mostly slop" output into "95%, ~senior level/sloppy staff level" output. You can tune the feedback loops so that the initial 80% pass gets better.

Don't just "add one more rule, bro" but look to establish deterministic pass/fail loops and let the LLM churn. You don't like the output because the methods are too complicated? Fail the tests if ABC score >25. Write the tests, disallow edit/write on them, and let it implement.

I think a lot of us are going to struggle to adopt this because the job is going to shift from "I can spend all day learning vi keybinds and yapping about OOP design principles" to "I am now a PM for an LLM". A lot of people are going to dislike this.

My advice to you if you feel that way is to find the things LLMs are consistently bad at (see up-thread) and move to specializing in those things. Because they are not bad at gluing together CRUD web apps, and 90% of us are paid to that.

If you're a builder first and a coder second, like me, you're probably very excited about this! You should be! You have now been given an indefatigable junior developer for $1000/month who will never complain! Go build things!

My final caveat is that all the old productivity metrics are now dead. PRs, commits, lines of code written all mean nothing. You cannot compare them to what you were doing before. Magical autocomplete can ship 10x the amount of PRS/commits/LOC you did before and it can be SHIT.

The measure of value is what is always has been: making stuff people like. Measure your productivity and the value of LLMs based on how quickly and how often you are doing that.