Back to Writing

Principles for Building Agentic & AI-Native Applications

Moving past chatbot wrappers: how streaming state, real-time voice, and agentic workflows change UI architecture.

The initial era of LLM integration was dominated by the conversational chatbox: a prompt input at the bottom, an auto-scrolling column of markdown text in the center. While chat is intuitive for exploratory questions, it is inadequate for complex software engineering and cognitive workflows.

Moving from Conversational to Agentic

True leverage comes when AI operates as an agentic partner embedded into context:

  1. Autonomous Tool Use with Human-in-the-loop: The agent executes deterministic actions (compiling code, executing tests, mutating filesystem state) while allowing the human to inspect and intervene.
  2. Multi-Modal Real-Time Latency: Low-latency voice pipelines (such as WebRTC or WebSocket audio streaming) enable spontaneous, conversational collaboration without the friction of typing.
  3. Deterministic Verification: Language models hallucinate; deterministic tool chains do not. Combining probabilistic generation with rigid compilers, linters, and test suites provides safety and confidence.

Architecture Patterns

When designing systems like Innata, the critical challenge is managing state synchronization between the client, the backend orchestration runtime, and third-party inference APIs.

“Simplicity of interface requires rigor of underlying state.”

Keeping the client state lean and relying on server-sent events or WebSockets ensures users experience instantaneous feedback without UI jitter.

QZ
Qiwen ZhaoAI Builder • Endurance Athlete • Stoic Thinker
Explore More Writing