Agentic software design is still unsettled: notes and tools to the rescue

Two threads today circle the same unresolved question about how to build software that works for both humans and AI agents. Hubble, an open-source notetaking app, is built with a dual interface from the start: a React UI for humans and plain markdown files for agents. The landing page critique was pointed: 'the burden of proof is high given how many of these there are.' The counterargument from a commenter: 'this is probably how a lot of new software projects will work.'

The LearnVector thread from Andrew Ng's new AI company raised the same tension from the product side. One commenter pushed back on calling AI-to-human tutoring 'one-to-one,' arguing the framing papers over a real qualitative difference. The edtech history of disappointing venture outcomes came up, alongside genuine optimism that AI changes the unit economics.

The pattern: the design language for agentic software is still being invented. Dual interfaces, plain files as the agent-readable layer, MCP as a protocol layer: none of these have won yet, and builders are experimenting in real time.