Great post Jonas, I agree with you that in 2026 we need to evolve our position with AI, maintaining a high-level view of operations without getting too deep into execution. Do you have a guide to set up the framework you outlined in this article?
I've been following The Circuit for the past year with great interest, and your insights have genuinely shaped how I think about agentic coding. So thank you for that.
The shift from static to fluid, modular workflows you describe really resonates. You mentioned custom N=1 software programs via MCP servers: for truly individual, ad-hoc automation cases, I'd argue there's an even leaner alternative worth considering: Claude Skills (quietly adopted by OpenAi) or similar approaches like Mario Zechner describes in his recent post "What if you don't need MCP at all?". The core argument: a smaller context window = less context rot
Yeah I use Claude Skills every now and then - they're great because they're portable across Anthropic environments. You can create one in Claude Code and use it in the Claude browser extension or on Claude.ai.
Coming from a software engineering background, they do have some key limitations: all code needs to fit in a single script file and they have a predefined runtime. So for more complex N=1 programs I tend to prefer MCP servers.
In addition, MCP is now an open source standard widely adopted in the industry so if I ever want to move away from Claude Code or the Anthropic ecosystem it will be more easy to move these workflow components elsewhere.
This modular approach to workflow design really nails the fundamental shift happening right now. The idea that workflows should be living organisms rather than fixed pipelines totally changes how we need to think about maintainance and iteration cycles. I've been experimenting with similiar patterns in my own automation stack, and the biggest win isn't even the cost reduction (though that's huge)... it's the psychological freedom from needing to get everything perfect upfront. When you can iterate a subsystem in minutes instead of hours, you start treating automation decisiosn as reversible experiments rather than permanent commitments, which ironically leads to way better design choices.
In a way this is good news for me. Finally I can stop doing most of the execution and lean on what comes natural which is high-level strategy, system design, and building workflows in code rather than n8n.
This post also highlights the fact that I need to carve some time to tinker around with MCPs and the different agentic models out there.
This is wild. AI isn’t just automating tasks anymore, it’s rethinking how we do work. The workflow itself becomes the product
Exactly!
Great post Jonas, I agree with you that in 2026 we need to evolve our position with AI, maintaining a high-level view of operations without getting too deep into execution. Do you have a guide to set up the framework you outlined in this article?
Thanks Ashwin!
I'm holding a live sessions on the topic this Thursday 5 pm CET on Cozora - link is in the post!
I'll also be updating my AI automation playbook for 2026 but that will be later this year.
Great piece, Jonas!
I've been following The Circuit for the past year with great interest, and your insights have genuinely shaped how I think about agentic coding. So thank you for that.
The shift from static to fluid, modular workflows you describe really resonates. You mentioned custom N=1 software programs via MCP servers: for truly individual, ad-hoc automation cases, I'd argue there's an even leaner alternative worth considering: Claude Skills (quietly adopted by OpenAi) or similar approaches like Mario Zechner describes in his recent post "What if you don't need MCP at all?". The core argument: a smaller context window = less context rot
Thanks Julius, much appreciated!
Yeah I use Claude Skills every now and then - they're great because they're portable across Anthropic environments. You can create one in Claude Code and use it in the Claude browser extension or on Claude.ai.
Coming from a software engineering background, they do have some key limitations: all code needs to fit in a single script file and they have a predefined runtime. So for more complex N=1 programs I tend to prefer MCP servers.
In addition, MCP is now an open source standard widely adopted in the industry so if I ever want to move away from Claude Code or the Anthropic ecosystem it will be more easy to move these workflow components elsewhere.
This modular approach to workflow design really nails the fundamental shift happening right now. The idea that workflows should be living organisms rather than fixed pipelines totally changes how we need to think about maintainance and iteration cycles. I've been experimenting with similiar patterns in my own automation stack, and the biggest win isn't even the cost reduction (though that's huge)... it's the psychological freedom from needing to get everything perfect upfront. When you can iterate a subsystem in minutes instead of hours, you start treating automation decisiosn as reversible experiments rather than permanent commitments, which ironically leads to way better design choices.
Unit economics ftw!
Systems thinking is finally where the leverage lives.
In a way this is good news for me. Finally I can stop doing most of the execution and lean on what comes natural which is high-level strategy, system design, and building workflows in code rather than n8n.
This post also highlights the fact that I need to carve some time to tinker around with MCPs and the different agentic models out there.
Thanks for laying it out so clearly. 👌