Data boundaries and human approvals for AI coding agents

By Rogier Muller10.30.26
Data boundaries and human approvals for AI coding agents

An agent should start a task already knowing what it may touch. We teach teams to write that boundary before the run, and to keep a person on every approval that leaves the repository. 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.

This is the same split as the methodology. How to review AI-generated code is where a reviewer notices the boundary was crossed. Evaluating AI coding agents as a team is the wider choice of which agent to trust with a class of work. This note is the boundary itself.

Name the files, commands and network first

Before the agent starts, write three limits for that task.

  • The files it may change.
  • The commands it may run.
  • Whether it may reach the network.

Put the limits in the task, not in a policy nobody opens. A missing test in one module is a boundary. "Improve the service" is not. If you cannot name the files, the task is not ready to delegate.

If the diff leaves the declared files, the change is not ready. Send it back with the boundary restated. Do not widen the boundary after the fact to match whatever the agent touched. Widening is a new task, with a new approval.

Architecture, security, production risk and business logic stay with a person. The agent can draft inside the files you named. It does not decide that a new path is in scope because the code was nearby.

Rolling this out in your team?

We train engineering teams on their own codebase in two days. Agent setup, shared rules and review included.

Record the boundary where the next person looks

A boundary that exists only in the chat disappears when the session ends. Record the decision in the repository, where the next person can read it without replaying the session.

AGENTS.md or a short project instruction file is enough when it says, for a class of task, which files and commands are allowed and who owns the merge. Keep it short enough that people will maintain it. A long policy that contradicts the last pull request is not a boundary. The pull request is.

Say what must be true before review: a diff, a test result or an explanation of why the test could not run, and a note of remaining risk. The reviewer uses that note to see whether the risky parts were checked. A confident summary with no file path is not the record.

Use the same written boundary for Claude Code, Cursor and Codex. The tools differ in where the agent runs. The files, the commands and the owner of the merge do not. Shared standards is how a team keeps that one rule across the three tools.

Approve each connection on its own

Model Context Protocol is an open protocol for connecting an AI application to external tools and context. Each connection needs a human approval. Prefer read access until a person has seen a useful result.

A server that can read issues is a different boundary from one that can change production data. Approve the connection for the task in front of you. Keep the payload small enough that the turn still holds the work. A connection the whole company can call, with no owner, is not an approval. It is an open door.

Read access is the default we start from in the room. Write access, and anything that changes a system outside the repository, waits until a person has seen a useful result and has named who accepts the next write. New dependencies need that same person. A library added because it was convenient is a connection to code you did not write, and someone has to accept it.

The team proficiency matrix calls the step after a shared workflow Governed: controls decide. You do not need a Governed programme to approve one connection. You do need a named person, a written allow-list for that task, and a habit of refusing the connection that was not on the list.

Keep a person on the merge

The agent may inspect and explain. The team owns the merge. That approval does not move because the boundary was written well. A clean boundary makes the review shorter. It does not make the review optional.

The person who approves the change owns requirements fit, design, and the verification of each finding. Automated comments can point at a line. They do not replace that decision. Security-sensitive code stays with a person even when the files were on the allow-list. Being allowed to read a file is not the same as being allowed to ship a change to it.

Delegate, Review, Own is the order. Delegate is the task inside the boundary. Review is the evidence on the pull request. Own is the person who can explain the result. Assisted, on the matrix, is still one person with a private helper. Delegated is an agent doing a whole task while a person reviews every change. Do not call a run Governed because a tool offered a toggle. Call it Governed when the written control is what decided.

Use the same rule in the workshop room

Delivery of the workshop follows the same rule we ask teams to write down. The session is under NDA. The exercises run in the team's environment, with tools the team has already approved. Code and data stay in the team's systems. We do not move the repository onto a machine the team has not accepted.

A standard workshop is two days. A narrower scope can be one day. A group is ten to thirty people, with a minimum of eight. Larger organisations train in successive groups of up to thirty. Delivery is onsite, offsite, or remote via Zoom. In the room, the boundary is practised on a real task: the files, the commands, the connection that is still read-only, and the person who will merge.

The first month of Pulse is included with a workshop. It is a shared Slack or Teams channel with both founders, a first response within two business days, a written briefing every two weeks, and a two-hour live session each month for up to ten engineers. Each briefing item carries one recommendation: adopt, test or ignore. A boundary that was too tight, or too loose, shows up there, on the next task, while the team still remembers why it was written.

Frequently asked questions

What is a data boundary for an AI coding agent?

The files, commands and external tools allowed for one task. If the diff leaves those files, the change is not ready. Send it back with the boundary restated.

Who approves a connection to another system?

A person. Each connection needs its own approval. Prefer read access until someone has seen a useful result. A server that can read issues is not the same decision as a server that can change production data.

Does the agent own the merge?

No. The approval on the merge stays with a person, and so does any step that writes outside the repository.

The engagement starts with a 30-minute introductory call, then the matrix, a stakeholder call with the owners of requirements and review, a written proposal, acceptance and scheduling, and bespoke training on your codebase. Book a 30-minute call when you want that scope.

Related research

Train your team on coding agents

Two days on your team's own repository, onsite or remote, with a written proposal before you commit.

Book a call

Prefer to measure first? Take the free team check.