Large migrations can combine machine-found locations, LLM edits, automated checks, and engineer review.
Software engineering · Global
Collections: Human still decides
Executive brief
The reusable pattern is not autonomous coding; it is AI generation inside a deterministic discovery, validation, and review envelope.
AI value · LLM contribution to accepted migration work
Verified74.45% of 595 submitted code changes and 69.46% of 93,574 edits were LLM-generated.
Experience report without a randomized comparator.
Before
Finds migration sites and writes context-specific edits across the monorepo. → Repairs failures and submits each change list.
After
Discovers likely change locations and prompts a monorepo-trained LLM to generate edits. → Reviews, repairs where needed, and submits validated change lists.
Human boundary
Automated checks may reject edits; engineers decide whether to repair and submit; reviewers retain merge approval.
Why it matters
Large migrations can combine machine-found locations, LLM edits, automated checks, and engineer review.
Before
How the work ran before the change.
Software engineer
Finds migration sites and writes context-specific edits across the monorepo.
Software engineer
Repairs failures and submits each change list.
What changed
Large migrations can combine machine-found locations, LLM edits, automated checks, and engineer review.
Decision rightHuman moves from creator to judge
After
How the same work runs now.
Migration tooling
Discovers likely change locations and prompts a monorepo-trained LLM to generate edits.
ControlLocation discovery, compilation, tests, and static checks.
Software engineer
Reviews, repairs where needed, and submits validated change lists.
ControlHuman review and standard submission controls.
Exception path
Failed or semantically difficult edits return to engineers for manual correction or conventional migration tooling.
Work removed
Decision authority
Automated checks may reject edits; engineers decide whether to repair and submit; reviewers retain merge approval.
Before
Software engineer
Finds migration sites and writes context-specific edits across the monorepo.
Control: Manual code review and testing.
Software engineer
Repairs failures and submits each change list.
Control: Reviewer approval.
After
Migration tooling
Discovers likely change locations and prompts a monorepo-trained LLM to generate edits.
Control: Location discovery, compilation, tests, and static checks.
Software engineer
Reviews, repairs where needed, and submits validated change lists.
Control: Human review and standard submission controls.
Work that left the path
Human role before
Engineers located and authored most migration edits manually.
Human role after
Three engineers supervised the migration, reviewed generated changes, handled difficult cases, and retained merge authority.
AI roleDecision mode: bounded generation. The LLM proposes context-sensitive code edits at discovered locations; it cannot merge code.
Traditionally manual, context-dependent migration edits.→74.45% of 595 submitted code changes and 69.46% of 93,574 edits were LLM-generated.
Twelve-month case study. · 39 migrations completed by three developers.
Experience report without a randomized comparator.
Earlier manual migrations.→Developers estimated a 50% reduction in total time.
Twelve-month case study. · 39 identifier migrations.
Time reduction is the developers' estimate, not instrumented time measurement.
Anti-pattern
Allowing model-generated repository changes to bypass compilation, tests, or human review.
Questions
Portability conditions
Reputation risk
low
Evidence and authority
Current · updated
1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 4a77ab1ce686bbd8
GitHub Copilot research experiment · Creator to judge
Inline generation can compress a bounded task but not prove lifecycle productivity.
Meta (Facebook) · Creator to judge
Routine static-analysis bugs can receive machine-proposed patches behind compilation and analyzer-clean gates.
Anonymous Fortune 500 software company · Creator to judge
AI can diffuse expert practice, with value concentrated among novices.
Accenture · Creator to judge
Developers can accept, edit, or reject inline suggestions while build, review, and merge controls remain.