Remove human checkpoints one at a time — each stage earned by ~10 consecutive green runs, and a single regression drops you one rung until you re-qualify.

Pattern

Progressive Autonomy Staging

The ladder

Four stages. Each removes one human checkpoint.

Stage1

Supervised

Human invokes /adw_plan_build manually

Human sees

every PR before merge

earned after ~10 consecutive green PRs

Stage2

Test-Gated

Trigger auto-runs plan + build on new issues

Human sees

PRs where tests failed

earned after ~10 consecutive test-pass runs

Stage3

Review-Gated

Trigger auto-runs plan + build + test + review

Human sees

review screenshots only on blockers

earned after ~10 consecutive clean reviews

Stage4

Zero-Touch Execution

adw_sdlc_zte auto-merges after full pipeline passes

Human sees

only rejected PRs

Each rung is a load test on the one below. Stage 4 doesn't get jumped to — it emerges after Stage 3 proves itself.

Markers to advance

Evidence, not intent. Count before you remove a checkpoint.

Stage 1 → 2

10consecutive green

PRs accepted

Proves The planner + builder produce merge-worthy code consistently.

Stage 2 → 3

10consecutive test-pass

runs, no intervention

Proves Test phase + auto-resolver catch what human review used to catch.

Stage 3 → 4

10consecutive clean

reviews, no blockers

Proves Review phase catches what human visual review used to catch.

"10" is a soft threshold — tune it to operator risk tolerance. The rule that matters: advancement requires evidence, not wishful thinking.

Rollback discipline

One regression drops you one stage.

Stage

4 · ZTE

12 clean runs · auto-merging

→

1 bad merge

Stage

3 · Review-Gated

counter reset · re-qualify

No papering over regressions. Staying at the same stage after a bad merge is how trust compounds negatively — quality degrades run by run and Stage 4 becomes Stage-4-in-name-only. The brake is what makes the ladder safe.

The rollback rule — not the advancement rule — is what prevents silent quality drift.

The anti-pattern

Skipping straight to ZTE destroys the ladder.

Skip to Stage 4unearned

Day 1 — enable ZTE everywhere

Day 2 — a few auto-merges land

Day 3 — first bad merge hits main

Trust vaporized — rollback to Stage 1

Months of potential progression, lost to one incident. The team never learns which link in the chain broke — because the chain was never tested link by link.

Staged rolloutearned

Stage 1 — 10 green · advance

Stage 2 — 10 green · advance

Stage 3 — regression · drop to S2

Re-qualify · advance · reach S4

Each link is load-tested before weight lands on it. Regressions are localized, diagnosable, and recoverable without collapsing trust in the whole pipeline.

Jumping to ZTE optimizes forward without a brake. The ladder is slower by design — and that's what lets it hold.

The discipline

Autonomy is earned by evidence, never granted by optimism.

Each stage proves the link before loading weight on it. Regressions drop you one rung — not zero — so trust compounds positively, link by link, all the way to zero-touch.