Been working on a way to get "agent-per-row" behavior in Postgres without actually running LLMs inside the database.
The problem: Calling LLMs from triggers/functions blocks transactions, exhausts connections, and breaks ACID. Saw some projects doing this and it felt dangerous for production.
The solution: DB-adjacent architecture. Lightweight triggers enqueue jobs to an outbox table. An external Python worker (agentd) polls, executes AI calls, and writes back safely with schema validation and CAS.
What you can build:
Auto-classify support tickets on INSERT
Content moderation that doesn't block your app
Lead scoring, fraud detection, and invoice extraction
Anything where data arrives and needs AI enrichment
Works with OpenAI, Anthropic, OpenRouter, or any Agent.
One SQL line to add AI to any table:
SELECT agent_runtime.agent_watch('tickets', 'id', 'classifier', 'v1', '{"priority":"$.priority"}');
Includes 9 example use cases in the repo. Would love feedback on the architecture.