Engineering AI maturity · a nine-stage model

Nine stages, and how
you know which one
you're in

Every stage below is a way of working, not a tool purchase. A short assessment places your team on the ladder; the full model sits behind it when you want the deep read. Honest answers make this useful. Optimistic ones make it decoration.

Reading this alongside the rest of the site

The ladder orders the other two things written up here. The Agentic PDLC is how a person works well in the loop. It earns its keep around stages five and six, where the spec is the artifact and a human steers every cycle. Loop & Harness Engineering is what you build when you'd rather be on the loop than in it: stages eight and nine, where the system does the judging and you watch the judging. The PDLC came first because that's the rung you climb first.

Find your stage · one team at a time

Nineteen questions at most, and two of them are gates that skip whatever doesn't apply to this team. All at one altitude: one team, one stream. Pick the closest answer, or say none fits. Each answer places part of the work on the ladder; the band your answers agree on is where the team sits.

These questions are for someone close to the work. Reading the org from the exec floor? The coherence check is the same axis at your altitude, measured from the scoreboard instead of memory.

The work New work usually arrives to this team as...
The work Who turns that ask into something a builder can start?
The work When a scope disagreement comes up, what settles it?
The work What does this team wait on most?
The work How often does this team's work reach production?
The build Do models write any of this team's production code?
The build Where does a model learn this team's rules?
The build Two builders ask the model for the same change. What comes back?
The build How much correction does a generated change need?
The build What stands between this team's changes and production?
The build Who reviews a change before it merges?
The system Does any of this team's work run start to finish with nobody driving?
The system When one agent hands work to another, what carries it?
The system Where do people sit in those runs?
The system What tunes the pipelines?
The system How does design work land in this team's build?
The system How does this team's finished work reach stakeholders and customers?
The system What does this team commit to at planning?
The system When this team's output gets better, what changed?
1 / 19
STAGE 1

Phase-gated delivery

Work moves through requirements, design, build, and test as sequential handoffs. The plan is fixed at the point of least knowledge and defended from there.

Waterfall Measure: Cycle time measured in months
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Governance that rewards plan adherence over shipped outcomes, and funding that releases against phase completion rather than working software.

The move to stage 2

Cut work into thin slices that each reach production. Replace phase sign-off with a demo of something running.

Arrow keys move between stages ·
STAGE 2

Iterative delivery

Work is sliced into increments that reach production on a cadence. What you learn inside the loop changes what gets built next.

Agile Measure: Deployment frequency · share of work shipped in under two weeks
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Iteration theatre: ceremonies that run on schedule while releases don't. Requests still arriving as prescribed solutions, so the loop learns nothing.

The move to stage 3

Put an assistant in the editor for everyone. Start measuring how much of a merged diff was drafted by a machine.

Arrow keys move between stages ·
STAGE 3

Assisted authoring

The editor proposes the next lines. Engineers type less and read more, but still own the structure of every change.

Copilot Measure: Suggestion acceptance rate · time to a first working draft
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Treating the tool as autocomplete and nothing more. No shared conventions in the repo, so every suggestion fights the codebase it lands in.

The move to stage 4

Move authorship out of the tab key and into a conversation. Ask for the change, then review the diff.

Arrow keys move between stages ·
STAGE 4

Prompt-driven development

Changes start as a stated intent rather than a keystroke. The craft shifts from writing code to describing the change and judging the diff.

Instruction first Measure: Share of merged diffs machine-drafted · review time per pull request
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

One-shot prompting with no supporting context, so quality is luck. Reviewers rubber-stamping generated diffs because reading them is harder than writing them was.

The move to stage 5

Make context durable. Commit the rules, patterns, and schemas the model needs into the repository where they're reviewed like code.

Arrow keys move between stages ·
STAGE 5

Context engineering

The knowledge a model needs is stored, versioned, and loaded on purpose. Prompts get shorter and output gets consistent across people.

Durable knowledge Measure: Variance in output quality between engineers · rework rate on generated changes
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Context rot: stale rules nobody prunes, and context that describes the codebase somebody wishes existed rather than the one in production.

The move to stage 6

Bring the model upstream of the code. Turn a problem statement into a reviewed specification before generation starts.

Arrow keys move between stages ·
STAGE 6

Spec-driven development

Engineer and model work out the specification together, then generate against it. Discovery collapses from weeks to hours because the spec is the thing you argue about and the thing you build from.

Intent as the artifact Measure: Time from problem statement to merged change · discovery as a share of cycle time
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Stakeholders who hand over solutions instead of problems. A culture that treats the spec as a document to file rather than the decision to fight about.

The move to stage 7

Let one scoped, reversible class of work run end to end without supervision. Choose the place where the blast radius is smallest.

Arrow keys move between stages ·
STAGE 7

Delegated execution

An agent takes a bounded unit of work from ticket to pull request unattended. Humans set the boundary and judge the result rather than driving each step.

Agents on bounded work Measure: Share of merged pull requests opened by agents · escape rate on agent-authored changes
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Autonomy without guardrails on blast radius, secrets, and protected data. Or the opposite: approval gates so heavy the agent saves nobody any time.

The move to stage 8

Chain the specialists. Give each handoff a contract, and put the human checkpoint where the risk actually concentrates.

Arrow keys move between stages ·
STAGE 8

Multi-agent orchestration

Specialized agents for spec, build, test, review, and compliance hand work to each other under an explicit protocol. Engineers design the pipeline instead of running the steps.

Pipelines of specialists Measure: Throughput per engineer · share of workflows completing without intervention
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Pressure is building to move up

What keeps teams here

Orchestration sprawl, with chains nobody owns. Handoffs that drop context, so the work degrades a little at every hop.

The move to stage 9

Close the loop. Feed review outcomes, defect data, and production telemetry back into the chains that produced them.

Arrow keys move between stages ·
STAGE 9

Compounding systems

The system improves from its own results. Review comments, escaped defects, and production telemetry change how future work gets generated without anyone rewriting a prompt.

Learning loops Measure: Quality trend at flat headcount · rate of improvement per unit of work
Your position
Not here yet

Nothing checked, or too few entry signals. If the stage below this one still describes the team better, work there first.

0 of 9 signals true
Entering Established Cresting

01 · Entering

The first signs you're operating this way

02 · Established

The stage is how the org actually runs

03 · Cresting

Past the edge of the map

What keeps teams here

Learning loops without evaluation, so the system drifts and nobody notices. Feedback that encodes reviewer taste instead of outcome quality.

Where it goes from here

There's no fixed next rung here. Keep the evaluations honest, keep the loop fed, and let the compounding do the work.

Arrow keys move between stages ·