Data boundaries and human approvals for AI coding agents

Set the data boundary before the agent starts, and keep a person on the approval. The boundary is the files, commands, and external tools allowed for one task. The approval is the merge, and any step that writes to a system outside the repository.

We teach this inside training because a faster agent with a wide boundary creates review work the team did not plan for. The methodology names the same split: delegate the task, review the evidence, own the result.

Write the boundary before the run

State the allowed files, the commands the agent may run, and whether it may reach the network. A small task from the real workflow is a better test than a feature list. Tighten one validation, update the failing test, and explain the diff.

Record the decision in the repository, not in a private chat. The next person should see what was allowed without replaying the session.

Evaluate AI coding agents as a team walks through that checklist.

Treat each external connection as its own approval

Model Context Protocol (MCP) is an open protocol for connecting AI applications to external tools and context. A server that can read issues is a different boundary from one that can change production data. Approve the connection for the task, and prefer read access until a person has seen a useful result.

The context window is the finite text an agent can hold in one turn. Treat it as a budget to spend deliberately. A tool that dumps a large payload into the turn makes the next steps worse, and it can pull data the task never needed.

Keep the human approval on the merge

The agent can draft a pull request and list the files it changed. A person still checks requirements, the diff, and the tests. Security-sensitive code stays with human review. New dependencies need a person to accept them.

If the agent left the declared files, the change is not ready. Send it back with the boundary restated.

Services is the ongoing version of this: a shared channel, a briefing every two weeks, and a monthly live session, or a workshop on your repository. Book a 30-minute call to set the first boundary with us.

Frequently asked questions

What is a data boundary for a coding agent?

The files, commands, and external tools the agent may use for one task, written down before the run starts.

What is MCP in this context?

Model Context Protocol is an open protocol for connecting AI applications to external tools and context. Each connection still needs a human approval.

Who approves the merge?

A person. The agent can draft the change and list what it touched. The team owns the merge.