Deeply insightful article, which I would love to see turned into an ongoing series.
While the article initially began as a bit hype-y (in the very human, non-LLM way!), I recognize that this sort of "hook" is what captures attention for the quick-scroll era of social information diet. I really enjoy the analogy to the music production revolution - it is an imperfect analogy, but a great illustration of both the transformational nature and the nuances on employment. It might just be me, but I did not appreciate the GarageBand effect as a layman / music enjoyer. Could this also be the future of software agents - most laymen do not realize its step function effect on a specific industry, until they notice an abundance of software offerings?
I applaud the section "The five pillars of AI code management", which gets into some specifics in how these tools affect both product and software cycles. There is a lot of hypothesizing based on few data points, such as BT Group and National Australia Bank measurements. I would love to see this notion of developer productivity empirically verified as coding agents are still very nascent as we sit here in July 2025. (some counter evidence for productivity here: https://x.com/metr_evals/status/1943360399220388093)
I realize there are two schools of thoughts:
1.) AI agents are all hype, pushed by the capitalistic structures incentivized to justify galactic valuations
2.) AI agents are the future and will introduce tech abundance and , and we're just in the first inning.
#1 focuses on current lack of evidence of massive productivity jump, and #2 is mostly speculative. Your article will likely receive a fair amount of criticism from followers of #1, and this is why I believe a series will be beneficial as models, agents, AND human management become sophisticated.
I would personally put this piece at the level of Andrej Karpathy's talk on Software 3.0 and Andrew Ng's BUILD 2024 Keynote! Thank you, @SidRao!
First of all, thank you for your kind words of encouragement. Andrew Ng and Andrej Karpathy are like gods, and I'm a mere mortal.
Second, you are correct about a series. I'm going to think deeply about how to structure this series. I'm currently not intelligent enough to jump to part two today, which is why it took me so long to reply to you. :-) I have an article brewing that starts with an Amadeus analogy. Salieri would probably be a follower of #1, bemoaning and pull his hair out over the output of an optimally constrained and enabled Claude Code thread.
"It's like I heard the voice of god through that damn CLI tool. But it also laughed at me at how quickly it output React pages of gold." -Salieri as a Coder in 2025
One thing, though - I have some emerging proof in the pudding. My open source project - Camille - demonstrates that Claude Code can, with guardrails, create productivity. I should compare the number of months it would have historically taken me to build Camille, and how I was instead able to get to v0.1.0 in a week.
But above all, I appreciate the encouragement. I sometimes wonder if I'm just writing for myself, and then I read reader feedback and I realize that my words still have an impact.
BTW - I have been discovering an entirely new way to leverage LLM agents—ad-hoc collaboration with competition and tools. I would love to get feedback on the approach... and collaborators: https://www.srao.blog/p/your-ai-needs-a-fight-club
Thank you for publishing this for me! I made a lot more friends!
Amazing article. I’m really inspired (after 25 years of software engineering)
Glad to hear!
Deeply insightful article, which I would love to see turned into an ongoing series.
While the article initially began as a bit hype-y (in the very human, non-LLM way!), I recognize that this sort of "hook" is what captures attention for the quick-scroll era of social information diet. I really enjoy the analogy to the music production revolution - it is an imperfect analogy, but a great illustration of both the transformational nature and the nuances on employment. It might just be me, but I did not appreciate the GarageBand effect as a layman / music enjoyer. Could this also be the future of software agents - most laymen do not realize its step function effect on a specific industry, until they notice an abundance of software offerings?
I applaud the section "The five pillars of AI code management", which gets into some specifics in how these tools affect both product and software cycles. There is a lot of hypothesizing based on few data points, such as BT Group and National Australia Bank measurements. I would love to see this notion of developer productivity empirically verified as coding agents are still very nascent as we sit here in July 2025. (some counter evidence for productivity here: https://x.com/metr_evals/status/1943360399220388093)
I realize there are two schools of thoughts:
1.) AI agents are all hype, pushed by the capitalistic structures incentivized to justify galactic valuations
2.) AI agents are the future and will introduce tech abundance and , and we're just in the first inning.
#1 focuses on current lack of evidence of massive productivity jump, and #2 is mostly speculative. Your article will likely receive a fair amount of criticism from followers of #1, and this is why I believe a series will be beneficial as models, agents, AND human management become sophisticated.
I would personally put this piece at the level of Andrej Karpathy's talk on Software 3.0 and Andrew Ng's BUILD 2024 Keynote! Thank you, @SidRao!
First of all, thank you for your kind words of encouragement. Andrew Ng and Andrej Karpathy are like gods, and I'm a mere mortal.
Second, you are correct about a series. I'm going to think deeply about how to structure this series. I'm currently not intelligent enough to jump to part two today, which is why it took me so long to reply to you. :-) I have an article brewing that starts with an Amadeus analogy. Salieri would probably be a follower of #1, bemoaning and pull his hair out over the output of an optimally constrained and enabled Claude Code thread.
"It's like I heard the voice of god through that damn CLI tool. But it also laughed at me at how quickly it output React pages of gold." -Salieri as a Coder in 2025
One thing, though - I have some emerging proof in the pudding. My open source project - Camille - demonstrates that Claude Code can, with guardrails, create productivity. I should compare the number of months it would have historically taken me to build Camille, and how I was instead able to get to v0.1.0 in a week.
Camille: https://www.github.com/srao-positron/camille
But above all, I appreciate the encouragement. I sometimes wonder if I'm just writing for myself, and then I read reader feedback and I realize that my words still have an impact.
Thank you!
Very inspiring, beautiful analogy — something random I wasn't expecting to open my eyes at this time of day (writing in CET -.-).
I'm surprised you didn't mention Gemini which is pretty good with now a CLI mode.
When I wrote the article, Gemini agent mode hadn’t been released. I didn’t have a chance to fully explore it. I will certainly take a look!
AI SLOP! this reads like it was written by grok during a power outage 🤭
Sorry dude. Not AI slop.
1) LLMs actually suck at coming up with analogies. Garage Band == SDEs, nope.
2) They also don’t understand the nuances of SDMs, SDEs, PMs, and leadership.
3) They also don’t make commentary about using their Claude API token when they run out of Max tokens.
4) Or talk about Charlie Bell who rarely did public talks before he went to Microsoft.
(I could keep going on)
But I did use Grammarly, which now has an LLM behind it. So - yes, AI did contribute to the creation of the article.
But if you are a gonna call that AI slop, you’re like one of the 80s record producers getting mad about iTunes and Napster. ;-)
Hugs and kisses.
sorry it just sounds like the manic ramblings of a desperate salesman to me. good luck 👍
Sorry it did not resonate with you! Good luck to you as well!
BTW - I have been discovering an entirely new way to leverage LLM agents—ad-hoc collaboration with competition and tools. I would love to get feedback on the approach... and collaborators: https://www.srao.blog/p/your-ai-needs-a-fight-club