AI-assisted coding
AI-assisted coding will not save a messy engineering org. It will make the mess faster.
Cursor and Copilot change the inner loop. They do not replace specs, review, or architecture. Here is the operating system we install around them.
Every engineering leader in Sydney has now seen a demo where a model writes a feature in four minutes. Some of those leaders have also seen the pull request that followed: tests that assert the mock, a secret in a fixture, and an architecture decision nobody remembers making.
AI-assisted coding is a multiplier. Multipliers do not choose the sign.
What actually changes
The inner loop gets cheaper: boilerplate, translations, test scaffolding, “what does this module do.” The outer loop does not automatically get cheaper: product spec, threat model, data contracts, operational ownership.
If you only accelerate the inner loop, you flood review. Seniors drown. Juniors look productive. The codebase accumulates confident wrongness.
The operating system we install
When we roll this into a team, we treat it as a delivery change, not a licence purchase.
- Specs as context. A one-pager in the repo beats a Slack thread the model cannot see.
- Golden paths. One way to make an API, a job, a React page. The agent copies what you already made excellent.
- Review that looks at diffs like a hostile compiler. Agents are prolific. Reviewers must be allowed to say “throw it away.”
- Eval the assistant. Keep a set of tasks you re-run when the tool changes. If quality drops, you will know before the board does.
- Data boundaries. Customer data, production dumps, and secrets do not go into a public model. Legal is not a vibe.
What we tell CISOs
Yes, the tools phone home. Yes, you need an allowlist of repos. Yes, you can still get a 2–4× inner-loop gain on the work that is well specified. No, you should not ban them and then wonder why your best people use them on personal machines.
The Australian version of this conversation includes Privacy Act, IP assignment, and “does this leave the country.” Have it once, write the policy, and stop having it every sprint.
The tell
You will know it is working when seniors spend less time typing and more time deciding. You will know it is failing when the backlog of “please review this 40-file PR” becomes the job.
