AI coding maturity: five levels and where your team stands

Ask five engineers in one team how they use AI for coding and you get five different answers. One delegates whole tasks to an agent, another still uses autocomplete, and the team has no shared picture of where it stands. The film above walks through the five levels of our team proficiency matrix in 60 seconds. This guide describes each level, how we measure a team, and what the result is for.

The five levels of AI coding

L1 Assisted. People write the code. AI autocompletes lines and answers questions in a chat. The way the team works has not changed yet.

L2 Delegated. An agent does a whole task, such as a bug fix or a small feature. A person reviews every change and checks the evidence: the diff, the tests, the result.

L3 Systematic. The team shares one way of working: the same instructions, context files and review checks in the repository. Anyone can pick up anyone else's work.

L4 Governed. Policies and automated checks decide what may merge. Agents run within set limits, every action is logged, and cost and quality are measured.

L5 Self-improving. The system improves itself: context, tests and workflows update from what happens in production. People set the direction and audit.

The levels describe a team's way of working. A single engineer can work at L2 inside a team that works at L1, and that gap is exactly what the measurement makes visible.

How we measure a team

1. Everyone on the team answers alone, in about six minutes. Engineers get 16 questions, product and design 12.
2. Each answer is a level from L1 to L5. A person's level is the average of their answers.
3. The shared baseline is the level three in four people have reached. Training starts there, so nobody is lost.
4. The target is one level up for the whole team, and one more for two or three workflows within 90 days.
5. You get a PDF assessment with the baseline, the advice and the matrix. It points to your next step.

Why the goal is one level up

A team that tries to jump from L1 to L4 skips the habits that make the higher levels safe. Delegation without review creates rework, and governance without a shared way of working has nothing to govern. Moving the whole team one level keeps everyone together and gives each step something concrete to practise on real work.

The baseline matters for the same reason. Training aimed at the most advanced engineer loses the rest of the team, and training aimed at the least experienced bores the others. Starting at the level three in four people have reached gives the group a common floor to build on.

See where your team stands

The team check is free and takes each person about six minutes. Every engineer answers alone, and you receive the team baseline and a recommended next step as a PDF. Start it at harness.institute/matrix.

Try this with your team: send the check to everyone who writes or reviews code, then pick one workflow the team runs every week and agree what one level up looks like for it.

Music in the film: "Signal to Noise" by Scott Buckley, released under CC-BY 4.0. www.scottbuckley.com.au