Skip to main content

Self-Improving AI Agents With Human Approval

SchemaBounce is where your business runs itself. It moves your data in, keeps your business knowledge and decisions, and runs agents next to them. Those agents get better at your business as they work, and they improve the platform they run on. A person approves the changes that matter.

This page explains how that works and where people stay in control.

What it means​

Three things live in one workspace:

  • Your data. Pipelines move data from your databases and SaaS tools into SchemaBounce.
  • Your knowledge and decisions. The Context Engine holds what your business knows. Decision records hold what your people decided and why.
  • Your agents. Hosted agents run on a schedule or on demand, with that data and those decisions within reach.

Because the agents sit next to the data and the decisions, they can see what worked and what didn't. They use that to build skills, watch goals, and propose changes to the product itself.

How it works​

Skills from their own runs​

When an agent finishes a run, it can turn what it learned into a reusable skill. The next run starts from that skill. Skills belong to your workspace and you can read them.

Standing goals​

A standing goal tells an agent which metric to watch. When the metric drifts, the agent opens work on the task board. The work carries the goal that triggered it, so you can trace any ticket back to the number that caused it.

Code sessions​

An agent can start a code session that changes the product. A person approves before the session starts, and its changes land on a branch your team reviews. When a person asked for the work and has a usable coding license, the session runs on that license and that person approves it. When no person owns the work, the session runs through the API on workspace credits.

Origin and approvals​

Every ticket records its origin: the person who asked for the work, or the schedule, webhook, or workflow that started it when no person asked. The platform sets the origin and copies it forward each time work passes from one agent to another. An agent can't change it. Risky actions wait in the Inbox until a person decides.

What stays with people​

  • Approvals. Actions marked as needing a human stop and wait in the Inbox. You set which actions those are in the workspace approval policy.
  • Decisions. An agent can propose a decision. Only a person with the right permission accepts or rejects it.
  • Code sessions. An agent's code session waits for a person's approval before it starts.

Agents don't ratify their own proposals, and they don't approve their own work.

Common questions​

Does SchemaBounce change things without asking?​

Agents act inside the permissions and approval policy you set. Actions that need a human wait in the Inbox until a person decides. A code session waits for a person's approval before it starts, and its changes land on a branch.

Who approves a code change?​

The person who asked for the work approves the session before it starts, and it runs on that person's coding license when they have a usable one. When no person owns the work, it runs through the API on workspace credits.

Who can accept a decision?​

Only a person with decision permission. Agents can propose decisions and cite accepted ones, but they can't accept or reject any.

Where do I see what an agent did and why?​

Each ticket records its origin: the person who asked, or the schedule, webhook, or workflow that started it. The Inbox shows pending approvals, and the task board shows the work an agent opened.

Ready to try it? See pricing or create a workspace.