Change management frameworks share an assumption so basic that it is rarely written down: a person wrote the change. Author, reviewer and approver are roles held by people, and separating them is what makes the control meaningful.
A coding agent breaks that assumption quietly, because the process still records people. The name on the commit is real. The approval is real. Neither describes who produced the code.
What actually changes
Authorship becomes ambiguous. The committer accepted the change; they may not have written a line of it. The control records them as author because there is no other field to use.
Segregation of duties gets thinner. If an agent wrote it and one person both accepted and approved it, the separation you are relying on is between a human and a tool that has no independent judgement.
Review depth becomes invisible. Reviewing two hundred lines somebody thought about is different from reviewing two hundred lines that appeared in four seconds, and your process records both as “approved”.
Restating the controls
You do not need a new framework. You need the existing one to describe what happened.
Record the actual author. Capture which agent and model produced each change, bound to the diff. The committer field stays; a provenance record sits alongside it.
Distinguish accepted from authored. Where a human revised the agent's output, record that separately. It is the strongest available evidence of meaningful review.
Require human sign-off where it matters. Not everywhere — that is how controls get bypassed. Name the directories: payments, authentication, anything a regulator opens first.
Report contribution per change. A pull request that is ninety percent AI-authored deserves different scrutiny from one that is five percent, and that is only actionable if the number exists at review time.
What not to do
Do not ban the tools. Adoption will continue without visibility, which is strictly worse than adoption with it.
Do not set a target for AI contribution. It is a fact to be known, not a number to optimise; the moment it becomes a target, people will route around the measurement.
Do not block on day one. A policy that stops commits before anyone trusts it gets switched off within a week. Warn, record, and tighten once the warnings are accurate.