The pipeline's job changes from catching mistakes to catching drift
A human-written pipeline mostly catches typos, missed edge cases, and the occasional logic error — the writer's intent was usually sound. An agent-written pipeline needs to catch something additional: silent drift from what was actually asked for, because an agent can produce plausible, well-formatted code that solves an adjacent problem to the one specified.
That means tests need to assert intent more explicitly than before — behavioural tests tied to the original requirement, not just unit tests tied to the implementation the agent happened to choose.
Code review becomes an audit of judgement, not of syntax
Reviewing an agent's diff for style and correctness is necessary but insufficient; the harder question is whether the approach taken reflects the actual trade-offs the task required, which an agent has no stake in getting right beyond what it was asked to optimise for.
Senior engineers increasingly spend review time asking "was this the right problem to solve this way" rather than "is this line correct" — a genuinely different skill, and one most review processes weren't built to make explicit.
Accountability doesn't move to the agent — it stays exactly where it was
Whoever merges the pull request is accountable for what ships, regardless of who or what authored the diff. Pipelines and review processes that quietly assume "the agent checked it" are building an accountability gap that only surfaces during an incident.
An agent can write the diff. It cannot be accountable for merging it. That line doesn't move no matter how good the model gets.MNB Software Engineering Practice
Key actions
- Add behavioural tests tied to intent, not just unit tests tied to implementation
- Review agent diffs for approach and trade-offs, not only syntax
- Keep merge accountability explicitly with a named human
- Track drift from the original requirement as its own review category