Building AI Native Software Beyond Agents
AI Engineergo watch the original →
the gist
Agents are the 'web pages' of the current era, serving as a primitive for a new category of AI-native software that will eventually replace them.
The Evolution of Computing Primitives
Computing history follows a trajectory of increasing abstraction, moving from the abacus to stored-program computers, personal computers, and the cloud. Each era introduced a new primitive that expanded the scope of human work rather than replacing it. VisiCalc serves as the primary historical analogy: it did not replace accountants but instead enabled vastly more accounting and created entirely new categories of work that were previously impossible. The current agent-centric development phase is a transitional primitive, similar to the static web pages of 1995, which eventually evolved into the complex web and mobile applications of today.
Building AI-Native Software
True AI-native software requires moving beyond simple agent loops toward multi-model harnesses that integrate tools, data, and context. The goal is to realize the visions of historical concepts like Apple's 1987 'Knowledge Navigator' using modern, real-world technology. The project 'Gradient Bang' demonstrates the next set of primitives required for this shift:
- Asynchronous non-blocking context compression: Managing state updates without stalling the primary interaction loop.
- Long-running subagents: Utilizing specialized agents (e.g., exploration or trade loops) that share context to perform background tasks.
- Progressive skills loading: Dynamically fetching or activating capabilities based on the current task requirements to maintain efficiency.
- Dynamic interface generation: Creating UI elements on the fly based on the agent's state and user needs.
- Conversational voice: Integrating real-time, low-latency audio as a primary interaction layer rather than a secondary feature.