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.
Related
- The context.md Pattern for Agent Memory — the mechanism behind accumulated context
- Compound Engineering - Only Document Mistakes — the same compounding principle applied to CLAUDE.md in Boris Cherny’s workflow
- Self-Modification Must Be Legible — the safety frame for the most advanced layer of improvement
- Composability - New Features Through New Prompts — composability is the substrate that lets prompt edits ship behavior change
- Three Modes of Working With AI — prompt refinement is the surface where the modes meet