ce-agent-native-architecture

Agent-native architecture treats agents as first-class citizens by mapping UI actions to atomic tools with explicit completion signals. It enables iterative loops of small composable tools and prompts to achieve outcomes while preserving parity with UI actions.

Map UI actions to atomic tools by creating small, composable primitives like read_file and write_file. Define outcomes in prompts rather than hard-coded workflows, and use explicit completion signals so agents declare when tasks are complete.

Dynamic capability discovery allows agents to identify available actions at runtime, while context injection feeds live app state into prompts. This ensures agents operate with current data while maintaining shared workspace parity between UI and agent operations.

Ensure parity by making the UI and agent operate in the same shared data space. Both should use identical atomic primitives and explicit completion signals, guaranteeing transparency and preventing data drift between manual and agent-driven workflows.

Test agent-native architecture by validating tool definitions, running agent loops with simple system prompts, and verifying explicit completion signals. Use example tool definitions to demonstrate the end-to-end workflow and confirm shared workspace parity.

Avoid prompt-driven agent architecture when workflows require strict deterministic execution rather than iterative loops. If your application cannot support dynamic capability discovery or shared workspace parity, hard-coded workflows may be more reliable.