So my Nano-Kitty side project is done with the first iteration. It's actually pretty good. The agent works, but connection management is nascent, and not what I would consider robust enough to be releasable.
It's just the reality of the way providers work. You want to be able to wire in redundancies, survive rate limiting and token limits, and potentially work with any and all models available when you're working on your happy little long haul projects.
When I'm done, this will work the same way it did in the Sugarfang prototype, only over a more heterogeneous network of devices within your "litter."
What's exciting here, is that it gives Spec Kitty a model agnostic autonomous agent, or an autonomous sub-agent that you can hook into an MCP and use from your favorite system.
What's interesting is the way it handles parallel missions on the same project. There were some changes around repo shadowing and parallelism that went into the main Spec Kitty repo last week that I'm taking full advantage of.
The concept is that in the worst-case scenario, the absolute worst thing that's going to happen is a fixit ticket. But Nano-kitty goes further, by actually isolating each work tree into a gated fully isolated process that disappears when it's finished running, which adds a layer of safety by fully isolating competing worktrees and memory spaces.
In review, the only real defect I found so far was that Codex decided to second guess me on the Agent's autonomy, and it re-inserted the human in the loop, which would have defeated the whole purpose of the app.
Outstanding:
- The MCP
- The Doctrine DAG.
When I'm done, you'll have an agent or a dedicated sub-agent that can run on any model, and complete some pretty extreme long horizon tasks.
The existence of the system is proof that it works, because it bootstrapped itself into existing.
It is its own integration test.
Pretty meta.
Hoping to drop code this week, or early next.
Yes yes.
More soon.
I love you all.