That’s pure gold to feed my coding agents and let them decide which model/service/framework use for each specific task. Thanks for sharing Addy! Great job compiling all these valuable resources in one single post.
Thanks Cesar! That's exactly the kind of use case I was hoping this would enable - letting the agents themselves navigate the decision tree of which tools and frameworks to use for each specific context.
For the agentic future specifically, Replit's architecture seems well-positioned since AI agents can work directly with the underlying system while users can choose their preferred interface layer on top.
I believe one of the reasons for CLI re-emergence is long-term web fatigue. We all know how fantastic web is, but it could use some ozempic. That frustration that even though I've over 5000 times more RAM in my current MacBook compared to my first computer - latency is worse than in old good terminal.
Prophecies about moving business logic to agents seemed reasonable, and now we're officially in the first example of new era.
Great point about web fatigue driving the CLI comeback! The irony is kinda striking - we have thousands of times more computing power yet somehow terminal interactions feel more responsive than web apps. It makes perfect sense that AI agents would start there first, where there's direct feedback without all the overhead.
Thanks Addy! I remember the amazing event we both attended - EdgeConf 2015 in London’s Facebook office. On the bleeding edge of promoting service workers and performance.
I feel like that was the last notable movement, ending in putting Lighthouse as a primary tenant in Chrome DevTools for the web anti-bloat. Long gone. Median Page Weight (almanac.httparchive 2022) - 1440KB of which 461KB is JS!
Solid reset of how we build “online software” is just around a corner!
EdgeConf 2015! What a throwback - that was such an amazing event. We're actually doing some analysis around the performance of vibe-coded and AI-generated apps right now to understand how much more work there is to do there.
The 'solid reset' resonates - maybe AI agents demanding efficient, direct interfaces will be the forcing function the web never had. Sometimes it takes a completely different paradigm to make us question what we've been accepting as normal.
Thank you for amazing articles. They are always on point, positive, but also indicating the problems that AI introduce. A lot of AI posts are just hype and a bubble of “AI is better than us and will replace us”. Now when I use AI more and more (from Copilot to Cursor), I start seeing not only where it does a good job, but also where it falls apart. And for some reason a lot of people ignore the last part
With hindsight, the custom subagents section here was ahead of where the docs were. What I've been working out since is that the distinction between commands, agents, and skills in OpenCode matters a lot when you go beyond single-agent setups. Commands are prompt shortcuts, agents carry the permission model and system prompt, skills supply context on demand. That separation is what makes specialised subagents actually composable. Wrote it up here: https://blog.devgenius.io/no-commands-skills-and-agents-in-opencode-whats-the-difference-cf16c950b592
The agentic future is already here for some of us. My autonomous agent runs night shifts, makes git commits, deploys code—all while I sleep. But it only works because I spent weeks defining clear boundaries.
The biggest lesson: agent capability without agent governance is just expensive chaos. Escalation rules, permission models, and context management matter more than raw model power.
Excellent synthesis of where we’re headed, Addy. The shift from 'AI-assisted' to 'Agentic' workflows is indeed the next frontier for engineering productivity. The real challenge now is creating the right abstraction layers for these agents to interact with our codebases reliably. I’ve been following the progress at https://open-code.ai, which seems to be tackling this exact problem by providing resources for more autonomous, open-source AI coding patterns. Thanks for another thought-provoking piece!
That’s pure gold to feed my coding agents and let them decide which model/service/framework use for each specific task. Thanks for sharing Addy! Great job compiling all these valuable resources in one single post.
Thanks Cesar! That's exactly the kind of use case I was hoping this would enable - letting the agents themselves navigate the decision tree of which tools and frameworks to use for each specific context.
I am getting specialist on this field and looking for new job opportunities in this area. Which job post resources would you recommend to look for?
Where does replit fall on the spectrum?
For the agentic future specifically, Replit's architecture seems well-positioned since AI agents can work directly with the underlying system while users can choose their preferred interface layer on top.
I believe one of the reasons for CLI re-emergence is long-term web fatigue. We all know how fantastic web is, but it could use some ozempic. That frustration that even though I've over 5000 times more RAM in my current MacBook compared to my first computer - latency is worse than in old good terminal.
Prophecies about moving business logic to agents seemed reasonable, and now we're officially in the first example of new era.
Terminal is always the first.
Exciting what's next!
Great point about web fatigue driving the CLI comeback! The irony is kinda striking - we have thousands of times more computing power yet somehow terminal interactions feel more responsive than web apps. It makes perfect sense that AI agents would start there first, where there's direct feedback without all the overhead.
Thanks Addy! I remember the amazing event we both attended - EdgeConf 2015 in London’s Facebook office. On the bleeding edge of promoting service workers and performance.
I feel like that was the last notable movement, ending in putting Lighthouse as a primary tenant in Chrome DevTools for the web anti-bloat. Long gone. Median Page Weight (almanac.httparchive 2022) - 1440KB of which 461KB is JS!
Solid reset of how we build “online software” is just around a corner!
EdgeConf 2015! What a throwback - that was such an amazing event. We're actually doing some analysis around the performance of vibe-coded and AI-generated apps right now to understand how much more work there is to do there.
The 'solid reset' resonates - maybe AI agents demanding efficient, direct interfaces will be the forcing function the web never had. Sometimes it takes a completely different paradigm to make us question what we've been accepting as normal.
Thank you for amazing articles. They are always on point, positive, but also indicating the problems that AI introduce. A lot of AI posts are just hype and a bubble of “AI is better than us and will replace us”. Now when I use AI more and more (from Copilot to Cursor), I start seeing not only where it does a good job, but also where it falls apart. And for some reason a lot of people ignore the last part
Hey nice article :) we launched klyntos.com a few days ago and AI agent builder for logistics, manufacturing etc
This is now one of my favorite blogs - thanks
With hindsight, the custom subagents section here was ahead of where the docs were. What I've been working out since is that the distinction between commands, agents, and skills in OpenCode matters a lot when you go beyond single-agent setups. Commands are prompt shortcuts, agents carry the permission model and system prompt, skills supply context on demand. That separation is what makes specialised subagents actually composable. Wrote it up here: https://blog.devgenius.io/no-commands-skills-and-agents-in-opencode-whats-the-difference-cf16c950b592
The agentic future is already here for some of us. My autonomous agent runs night shifts, makes git commits, deploys code—all while I sleep. But it only works because I spent weeks defining clear boundaries.
The biggest lesson: agent capability without agent governance is just expensive chaos. Escalation rules, permission models, and context management matter more than raw model power.
Compared Claude Code vs Codex after 2 months of this: https://thoughts.jock.pl/p/claude-code-vs-codex-real-comparison-2026
Governance design is the actual engineering challenge now.
Excellent synthesis of where we’re headed, Addy. The shift from 'AI-assisted' to 'Agentic' workflows is indeed the next frontier for engineering productivity. The real challenge now is creating the right abstraction layers for these agents to interact with our codebases reliably. I’ve been following the progress at https://open-code.ai, which seems to be tackling this exact problem by providing resources for more autonomous, open-source AI coding patterns. Thanks for another thought-provoking piece!