Good question. Why hasn't there been a profusion of new game-changing software, fixes to long-standing issues in open-source software, any nontrivial shipped product at all? Heck, why isn't there a cornucopia of new apps, even trivial ones? Where is all the shovelware [0]? Previous HN discussion here [1].

Don't get me wrong, AI is at least as game-changing for programming as StackOverflow and Google were back in the day. I use it every day, and it's saved me hours of work for certain specific tasks [2]. But it's simply not a massive 10x force multiplier that some might lead you to believe.

I'll start believing when maintainers of complex, actively developed, and widely used open-source projects (e.g. ffmpeg, curl, openssh, sqlite) start raving about a massive uptick in positive contributions, pointing to a concrete influx of high-quality AI-assisted commits.

[0] https://mikelovesrobots.substack.com/p/wheres-the-shovelware...

[1] https://news.ycombinator.com/item?id=45120517

[2] https://news.ycombinator.com/item?id=45511128

I use GitHub Copilot in Intellij with Claude Sonnet and the plan mode to implement complete features without me having to code anything.

I see it as a competent software developer but one that doesn't know the code base.

I will break down the tasks to the same size as if I was implementing it. But instead of doing it myself, I roughly describe the task on a technical level (and add relevant classes to the context) and it will ask me clarifying questions. After 2-3 rounds the plan usually looks good and I let it implement the task.

This method works exceptionally well and usually I don't have to change anything.

For me this method allows me to focus on the architecture and overall structure and delegate the plumbing to Copilot.

It is usually faster than if I had to implement it and the code is of good quality.

The game changer for me was plan mode. Before it, with agent mode it was hit or miss because it forced me to one shot the prompt or get inaccurate results.

Here's one - https://apps.apple.com/us/app/pistepal/id6754510927

The app is definitely still a bit rough around the edges but it was developed in breakneck speed over the last few months - I've probably seen an overall 5x acceleration over pre-agentic development speed.

I use Junie to get tasks done all the time. For instance I had two navigation bars in an application which had different styling and I told it make the second one look like the first and... it made a really nice patch. Also if I don't understand how to use some open source dependency I check the project out and ask Junie questions about it like "How do I do X?" or "How does setting prop Y have the effect of Z?" and frequently I get the right answer right away. Sometimes I describe a bug in my code and ask if it can figure it out and often it does, ask for a fix and often get great results.

I have a React application where the testing situation is FUBAR, we are stuck on an old version of React where tests like enzyme that really run react are unworkable because the test framework can never know that React is done rendering -- working with Junie I developed a style of true unit tests for class components (still got 'em) that tests tricky methods in isolation. I have a test file which is well documented explaining the situation around tests and ask "Can we make some tests for A like the tests in B.test.js, how would you do that?" and if I like the plan I say "make it so!" and it does... frankly I would not be writing tests if I didn't have that help. It would also be possible to mock useState() and company and might do that someday... It doesn't bother me so much that the tests are too tightly coupled because I can tell Junie to fix or replace the tests if I run into trouble.

For me the key things are: (1) understanding from a project management perspective how to cut out little tasks and questions, (2) understanding enough coding to know if it is on the right track (my non-technical boss has tried vibe coding and gets nowhere), (3) accepting that it works sometimes and sometimes it doesn't, and (4) recognizing context poisoning -- sometimes you ask it to do something and it gets it 95% right and you can tell it to fix the last bit and it is golden, other times it argues or goes in circles or introduces bugs faster than it fixes them and as quickly as you can you recognize that is going on and start a new session and mix up your approach.

I know of many experienced and capable engineers working on complex stuff who are driving basically all their development through agents. This includes production level work. This is the norm now in the SV startup world at least.

You don't just YOLO it. You do extensive planning when features are complex, and you review output carefully.

The thing is, if the agent isn't getting it to the point where you feel like you might need to drop down and edit manually, agents are now good enough to do those same "manual edits" with nearly 100% reliability if you are specific enough about what you want to do. Instead of "build me x, y, z", you can tell it to rename variables, restructure functions, write specific tests, move files around, and so on.

So the question isn't so much whether to use an agent or edit code manually—it's what level of detail you work at with the agent. There are still times where it's easier to do things manually, but you never really need to.