Improvement Over Time Without Shipping Code

The Idea

Unlike traditional software, agent-native applications can improve without shipping code. The system gets better through three mechanisms:

  • Accumulated context. State persists across sessions via context files. The agent learns what exists, what the user has done, and what worked, and uses that on every subsequent run.
  • Developer-level prompt refinement. The team ships updated prompts to all users without a release. Changing the agent’s behavior is a config edit, not a deploy.
  • User-level customization. Users modify prompts to fit their own workflow without waiting on the product team.

A more advanced (and still emerging) layer is self-modification: agents that edit their own prompts or code based on feedback. This requires safety rails — approval gates, checkpoints, rollback paths, health checks — because behavior can drift in ways that aren’t immediately visible.

Why It Matters

Traditional software improves on the cadence of release cycles. Agent-native software improves on the cadence of conversations. The compounding is different in kind: every interaction can leave the system better tuned for the next one.