A swarm of agents searching for the best placing and routing of a PCB. Okay these are just two pads to be connected together. And they got closer and avoided a corner. Will test on more complex things.

Sometimes I stop and realise how much my life has changed in the past 12 months, after a lifetime of programming-as-a-lifestyle-and-identity. And I think we should be collectively be amazed at the capacity of us humans to adapt. Do I miss the old world? Yes, very much.

Well, I think when your agents start routing pcie lanes -- albeit a bit too corageously -- and pass DRC that you can finally call yourself impressed. (This is an existing PI hat, routing removed, then added again)

I might be having way more fun with this than i should. I am not convinced that roles help, but the do certainly make to process looks entertaining

After ASYS-SWARM, another one, both hilarious and useful: ASYS-SENATE. Modeled after the Roman Senate. Agents debate until consensus, or until the princeps senatus calls it. It can substitute a single agent in asys. As reviewers they’re deeply insightful — and accidentally comic.

I learned a new word on X: stigmergy. Implemented it in asys, for RTL search. 16 agents on an 8×8 grid, radius 1. Each one only sees published designs in that neighborhood, then writes its own. Optimising UART for LUT count on iCE40: 220 → 177, 16 rounds.

Okay, things _are_ getting interesting, despite the terrible job at placement and routing: agents produced the first go at an evaluation board for Erbium, the upcoming @Ainekk0 chip. Throwaway experiment for now, but better than a blinking light I had yesterday.

This is all very silly, and of course creating a simple 555 with a blinking led powered by USB-C is not hard, just useless. But part of this game was using THT to stress the agent's place and route of the pcb, using KiCAD. On bigger boards it was immediate. Here I had to tell to take a screenshot of the sheet and decide how to use the space. Amazingly, it worked. In any case, there you go, the first PCB generated by my new and upcoming agentic system. God knows what that pad/large via is doing there.

Oh, I miss my first computer. Yet another slot in my personal project queues.

I remember the project I learned verilog for. It took about six months of sleepless nights. Writing and rewriting the same thing, learning a lot (as usual) from my favorite kind of friends: the ones smarter than me -- low bar. In the end, it did what I wanted, fit into my super small FPGA, ICE40HX4K, and felt like a monumental task. Came back at it tonight. I was pleased that Astra found two bugs only, and both silly one-line fixes. But also realised that then it just didn't stop at whatever extreme requests I asked it to do. And all the stuff I was dreaming of implementing in the distant future are now commits in my hard drive. Essentially, personal projects before 2025 feel like remains of a glorious, past civilisation.

Your weekly coding agent tip: "Strongly prefer English words with latin and greek roots rather than germanic ones" improves massively the communication style and clarity.

Asking someone what they think and getting a summary of their SoTA model of choice back is exhausting. I don’t particularly like talking to humans, but I liked talking to engineers. They used to have opinions of their own.

Forget coding. Using LLMs to learn how to improve things in your expensive yet cheaply built 2010s British flat is the real revolution.

Hello Differential Analyser, you are my next project. The two disc integrators moved by meccano carts are a thing of beauty.

finally a better way to explain what cyclo is.

okay, an isometric viewer for one of my component systems has been the most satisfying use of claude code recently. About component systems, the idea of Programming in the Large while leaving Programming in the Small to agents is the way I am inclined to go.