Docs
Agent-Native Usage

Agent-native usage

Why Silos suits AI agents — databases that provision and tear down through the API as fast as an automated workflow needs.

Silos is designed to be agent-native: AI agents and automated workflows can create databases, do work against them, and destroy them entirely through the API, as fast as they need to. The same properties that make Silos good for serverless apps — instant provisioning, scale-to-zero economics, and per-database isolation — are exactly what an agent wants from a database.

Why agents and Silos fit

Provision in seconds

A database is metadata plus storage until queried, so creating one is fast — no server to wait on.

Destroy just as fast

Tear a database down through the API when the task is done. The lifecycle is symmetric.

Pay only for active work

Mostly-idle databases scale to zero, so many short-lived databases cost far less than many always-on servers.

Isolated by default

Each database runs in its own sandboxed runtime with storage scoped to it — one task can't reach another's data.

Real Postgres

Agents get full SQL — transactions, JSON/JSONB, RLS — not a limited key-value store.

API-first

Everything an agent needs is a REST call: create, get a connection string, run SQL, delete.

The core pattern

Agent workloads follow a simple, repeatable shape:

Create

The agent calls the API to provision a database and receives its identifier and connection details.

Use

The agent connects with the connection string and does its work — runs SQL, stores intermediate state, queries results.

Destroy

When the task finishes, the agent deletes the database (or lets a TTL reap it), and billing for it stops.

See Ephemeral databases for the create→use→destroy flow in full, and Lifecycle for TTL, cleanup, and cost considerations.

Where this shows up

  • Per-task scratch databases. Give each agent run its own isolated database for intermediate work, then discard it.
  • Per-tenant databases. Provision a database the first time a tenant appears and tear it down when they leave — each one bills only while active.
  • Preview and evaluation environments. Spin a branch or database for a trial, evaluate, and clean up.

On this page