X
@every: Agent-Native Architectures: How to Build Apps After Code Ends Software…
@everybookmarkx
TL;DR. Stop writing features as code—describe outcomes, give agents atomic tools plus filesystem access, and let prompts compose new capabilities you never built.
Takeaways
- Five principles: parity (agent can do anything the UI can), granularity (atomic tools, not bundled logic), composability (new features = new prompts), emergent capability (users do things you didn't anticipate), and improvement over time (context.md + refined prompts).
- Files are the universal agent interface: already-known primitives (cat/grep/mv), inspectable, portable, syncable, self-documenting—if a human can navigate your structure, an agent probably can too.
- The test: describe an outcome you didn't explicitly build a feature for. If the agent can loop its way to it using your tools, the architecture is working.