Shopify’s CEO described Pi as “the most interesting agent harness” — a tiny core that writes plugins for itself as you use it, reinforcement-learning itself into the agent you need. When he wanted Claude Code’s tasks feature, he told Pi to spawn Claude in tmux, interrogate it about the feature, and implement a version. The agent used another AI to learn, then modified itself.
This is the logical endpoint of In agent-native architecture, features are prompts — not code: features aren’t just described in natural language — they’re discovered by the tool observing what you need and self-generating. It inverts traditional software where you adapt your workflow to the tool. The AI's self-improvement loop means each generation builds the next one faster that Shumer describes at the model level is happening at the application level too — each self-extension makes the tool more capable, which surfaces more extension opportunities. The risk is that improving models will eat this scaffolding: self-generated plugins are ephemeral, and better models may internalize them entirely. Shipper takes this further with Personal software grows through relationship, not configuration — the software doesn’t just extend itself mechanically, it develops a personality through the relationship with its user.