Agent-Native Infrastructure

Agents work better when the surrounding software, documentation, and operational tooling are legible to them from the start.

Summary

Karpathy argues that much of today’s infrastructure is still written for humans. Documentation tells the human to click through URLs, configure settings, and wire services together. Agent-native infrastructure would instead expose instructions, sensors, actuators, APIs, and data structures that are legible and executable by agents.

Examples

• Setup instructions written as agent-ready tasks.
• APIs and CLIs that avoid fragile browser-only configuration.
• Docs that include copy-pasteable agent prompts.
• Machine-readable project state and deployment state.
• Agents that can coordinate with other agents on behalf of people or organizations.

Implication For This Wiki

This wiki should be easy for an LLM to operate:

• notes/index.md tells the agent what exists.
• log.md tells the agent what happened recently.
• AGENTS.md tells the agent how to act.
• raw/Source Index.md tells the agent what sources exist and their status.

That makes the vault more agent-native than a normal folder of notes.

Related Concepts

• Agentic Engineering
• Software 3.0
• LLM Knowledge Systems

Sources

• Andrej Karpathy: From Vibe Coding to Agentic Engineering

Open Questions

• What would make this Obsidian vault more agent-native?
• Which operations should be represented as CLI tools instead of prose instructions?