https://github.com/mistercrunch/db-agents

DB-AGENTS is a proposal for an AGENTS.md-style convention that embeds agent-readable semantics and guidance directly inside databases.

This document describes a lightweight convention, not a platform or framework.

DB-AGENTS enables AI agents to better interface with databases by exposing contextual guidance inside the database itself. It requires no specialized tooling: adoption can start with a single metadata table and a small set of instructions injected into a system prompt.

Motivation

Modern AI agents are already excellent at writing SQL.

What they lack is not capability, but context:

• Which tables are canonical?
• Which joins are safe or dangerous?
• What definitions matter?
• What constraints must be respected?
• What conventions are expected?

Traditional semantic layers attempt to solve this through rigid schemas, metrics DSLs, and enforced abstractions. That rigidity made sense for deterministic BI tools—but it is increasingly misaligned with agentic systems, which reason probabilistically, adaptively, and incrementally.

DB-AGENTS proposes a different approach:

A soft semantic layer, expressed as documentation, stored inside the database, and discoverable by agents at query time.

Inspiration: AGENTS.md

DB-AGENTS is inspired by how AGENTS.md files are used in code repositories.