The New Primitives: Building AI Native Software — Kwindla Kramer, Daily
The talk argues that agents are an early interface, not the endpoint, and demonstrates reusable primitives for AI-native software: persistent subagents, compressed context, loaded skills, generated UI, and voice.
Kwindla Kramer frames today’s agents as a transition toward software built around inference. His multiplayer demo uses **hundreds of inference calls** and shows long-running subagents, shared context, non-blocking context compression, generated interfaces, and voice.
Treat the harness as the product architecture: decide how tasks persist, how context is compressed and shared, when skills load, and how interfaces adapt. The demo suggests these primitives can support interactions beyond a chat box or single agent loop.
Kwindla Kramer frames today’s agents as a transition toward software built around inference. His multiplayer demo uses **hundreds of inference calls** and shows long-running subagents, shared context, non-blocking context compression, generated interfaces, and voice. Treat the harness as the product architecture: decide how tasks persist, how context is compressed and shared, when skills load, and how interfaces adapt. The demo suggests these primitives can support interactions beyond a chat box or single agent loop. The argument is exploratory, not an evaluation. The game demonstrates that these patterns can run together, but provides **no benchmark, cost breakdown, or reliability data** for choosing them in production.
This broadens harness engineering from coding workflows into the runtime architecture of inference-native software: persistent tasks, cooperating subagents, shared and compressed context, generated interfaces, and voice become product primitives. It reinforces progressive skill loading and deliberate compaction, while the demo-only evidence leaves production cost, reliability, and architecture choices unresolved.