Git for Cognition
We consider an agent a few dozen turns deep, forked a hundred ways from one decision point. Which version-control verbs survive contact with a mind ... and which ones don't.…
}
Skip to main content9 posts
We consider an agent a few dozen turns deep, forked a hundred ways from one decision point. Which version-control verbs survive contact with a mind ... and which ones don't.…
Current agents can remember. They cannot inherit themselves. The third post in a series on agent architecture argues that the missing layer isn't better memory — it's a persistent, evaluable environment where an agent's past work remains live, callable, and composable.…
Current agent frameworks treat tool definitions as inert data the model can't reason about. The fix isn't a better format, it's a representational property called homoiconicity, and it changes what agents are capable of becoming.…
Today's AI agents can call tools, but ... they cannot evolve them. The bottleneck isn't reasoning or context length. It's representational: actions, traces, and plans live in incompatible formats that prevent agents from inspecting or composing their own capabilities…
(Or, how the "Solitude Curve" actually works) The solitude curve describes the pattern: AI assistance lets you work independently for longer stretches before needing human intervention. But…
Whether we're in 1957 looking at the first compiler, or 2025 looking at AI assistants, the dynamic is similar: automation that extends how long you can work productively without external dependencies.…
... maybe ... we're watching the end of the era where you could meaningfully opt out of big tech platforms while remaining professionally competitive. Maybe AI integration requirements make that stance unsustainable for anyone who wants to participate fully in modern work.…
I can write a fuller post on this later, but here's a sneak peek at how things can be, when you let ClaudeCode be "a local delegate" to your Vault. Here's it processing my Inbox for…