Project update:
The coding agent project is coming along.
I’m still working through the Pizza Empire benchmark. We just passed 17 of 29 tickets, which is the furthest I’ve ever gotten with it.
The swarm is working. The new diffusion process is in and fully functional. The new memory upgrades are *flawless.* The planning graph is doing what planning graphs do, and QA is being its usual professionally dignified self.
This morning, I borrowed some planner workflow ideas from my friend Robert Douglass, who was nice enough to remind me that his agent system, Spec-Kitty, still exists.
I took a quick look at it and realized he had written a better core planning loop than I had, so I did the honorable thing and stole it.
The refactor was absolutely massive. There is a lot in this system that depends on the planning loop, but it's so much smoother than it was yesterday.
Eventually, we hit a failure in the backlog orchestration when we realized the planner had created alternate timelines inside the project: two conflicting, contradictory tickets that caused the whole benchmark to melt.
But the good news is that we learned something.
All this new planner needs is a little bit of backlog validation before the implementation loop starts.
Correct the orchestration... finish the benchmark.
So that’s what I’m building now.
More soon 🕯️