Rui CarmoI’ve had some feedback that my last few weekly notes (especially over the holidays) have been a bit too long and that I should try to keep them shorter and more focused on a single topic.
Well, this week I wrote a lot thanks to the inclement weather and some insomnia, so I broke most of it out into separate posts:
- Seizing The Means Of Production (Again) covers my ongoing efforts to build a minimal-friction agent management setup using
agentboxandwebterm. - Vibing with the Agent Control Protocol dives into my new Slack-like web interface for chatting with AI agents via ACP, called
vibes. - Thoughts on AI-Assisted Software Development in 2026 is a quick brain dump of some of the lessons I’ve “learned” this week.
This is the age of TikTok, after all, and attention spans are shorter than ever, so I might as well try to adapt.
But there was a lot more going on this week, so here’s a quick roundup of two other things I worked on:
go-rdp Improvements
My web-based RDP client got UDP transport support—experimental, gated behind a --udp flag, but apparently functional. So far it’s going well, even if MP3 audio support is not all there yet (and yes, I know it’s a bit much for a web client).
This is one of my “things that should exist” projects that I’ve been using to experiment how to spoon-feed AI agents with highly complex protocol specs, and it’s been working out great so far, largely because RDP is so well documented.
The test suite now includes Microsoft Protocol Test Suite validation tests for RDPEUDP, RDPEMT, RDPEDISP, and RDPRFX. It’s the kind of spec compliance work that’s tedious but essential, and that I am increasingly convinced can be massively sped up with AI (after all, it took me less than a working week of continuous effort to do the whole thing, instead of the months it would have taken me otherwise).
go-ooxml from OO to… Hero?
As another data point, I decided to build my own Office Open XML library in Go this week, called go-ooxml. It too is meant to be a clean-room implementation, and I intend it to be a comprehensive replacement for the various format-specific Python libraries out there–I’ve already been using Go as an accelerator for Python in things like pysdfCAD, so I know it works.
And it also fits into the “things that I think should exist” category, because existing Go libraries for this are either abandoned, incomplete, or have APIs that make me want to cry.
And on the AI front the ECMA-376 spec is perfect:
- It is a real thing people used for other implementations, and progress has been ridiculous–agents are genuinely useful for this kind of spec-driven implementation work.
- In about two days (less than four hours of “actual work”), I went from initial project setup to enough to handle basic document creation (mostly Word, which is the most complex)
Is it production-ready? Of course not.
But it’s already more complete than many alternatives, and having the foundation right means I can add features as I need them without fighting the architecture.
Agents make the breadth possible, but they don’t remove the need for taste: deciding what the public API should look like, what’s testable, and what’s going to be maintainable six months from now is, of course, the hard part still to come…
At least I’m aware of it, I guess.