Post

Yegge is certainly a flavour. This so ahead of the curve I'm considering it as a piece of speculative design. Interesting: 1. Design and planning becomes the bottleneck when you hand all production code over to agentic systems. Design is a bottleneck *even when* the interface is just a TUI...

2. The cost could be as much as $2k+ / month? (wild speculation - open to more accurate guesses) At this early stage the efficiency of the system is literal shit, so the cost ≈ valuable output equation is inaccurate. But companies would pay ~near this for a high quality, low waste version of this

3. The conceptual design here is all over the shop – polecats, deacons, dogs, boot the dog, seancing, beads, epics, molecules, protomolecules, MEOW stack, convoys, patrols, refinery, workflows. I got lost somewhere after the MEOW stack... But these point at the shape of future agentic primitives:

There's a central orchestrator, implementation monkeys, a git merge resolver, an unsticker who troubleshoots issues, a work hustler, maintenance workers, special ops teams. Agents with specific roles working in a cohesive flow with trad programming supports: git, daemons, issue tracking, tmux

I'm not going to claim I entirely understand the workflow primitives even after a couple reads (does this diagram help? no? no) But the gist seems to be breaking tasks down into atomic units of work, tracked in git, executed, checked, and merged through a zoo of loops and pre-defined workflows.

0 reposts6 likes

I'm in fruitful debates with colleagues over projects like this. Namely: how much hands-on access to code do we think developers will need in the next generation of tools? How close does it need to be? I'm in the code-must-be-close camp, at least for the next 2 years. My bias is I do front-end...

...and language is a poor medium for designing easing curves and aesthetic feelings. I always need to touch the code, and it's often faster to CSS than prompt. FE requires constant human-in-the-loop checks so even if you have a dozen agents working, I am the bottleneck. Scale isn't that helpful.

I loved reading this. I like and respect the people behind it. They've done awesome stuff. I'm a happy user of GitHub Copilot's Agent mode. Do I recognise and am I convinced by this yet? No. But maybe I will be. It made me want to try building something directly with Claude Code to learn more

It's so hard to keep abreast of what's possible with models at the moment. I definitely suggest playing with Claude Code, Codex, & Gemini to develop your own felt sense of where we're at. Even just toy projects or personal benchmarks. They'll both surprise you and face plant. But in interesting ways