Agent-Native Architecture

Working rules

• Keep authority, scope, approval, and runtime isolation in trusted orchestration; prompts alone cannot enforce them.
• Provide explicit completion and partial-progress signals, durable state, and observable action results.
• Validate capabilities with real tasks, including failure and interruption paths; do not infer improvement from elapsed usage.
• Evaluate discovery from original inputs. Replaying fixed candidates measures classification and cannot reveal candidates a changed search would miss or invent. Compare each discovery variant with a fresh baseline on the same complete inputs in isolation. Independently adjudicate unmatched outputs from the complete result sets, including outputs from clean tasks.
• Calibrate decision gates per model. A replacement model's confidence scores do not preserve the previous model's cutoff, even when ranking accuracy improves. Fit thresholds on a separate calibration set. Verify false-positive and false-negative behavior through the production decision path on untouched representative cases.
• Use the selected topic's references and the architecture checklist to produce a concrete design with evidence and unresolved constraints.

Core Principles

Five principles govern agent-native design. For detailed explanations, examples, and test criteria, see core-principles.md.